full fua testing for simplify_network()

[jGaboardi]

This MR:

• resolves full FUA(s) for full algo run #38
• xref start filling up testing suite & docstrings #20
• xref start testing + refactor [stub issue] #21

[jGaboardi]

Failures we are seeing are unfortunately not due to small differences in testing precision, but actual _status difference. For example, in Aleppo:

>       assert_series_equal(known._status, observed._status)

sgeop/tests/test_simplify.py:44: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
testing.pyx:55: in pandas._libs.testing.assert_almost_equal
    ???
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

>   ???
E   AssertionError: Series.index are different
E   
E   Series.index values are different (0.02026 %)
E   [left]:  Index([    0.0,     1.0,     2.0,     3.0,     4.0,     5.0,     6.0,     7.0,
E              8.0,     9.0,
E          ...
E          39658.0, 39659.0, 39660.0, 39666.0, 39668.0, 39669.0, 39670.0, 39676.0,
E          39678.0, 39679.0],
E         dtype='float64', length=39489)
E   [right]: Index([    0.0,     1.0,     2.0,     3.0,     4.0,     5.0,     6.0,     7.0,
E              8.0,     9.0,
E          ...
E          39658.0, 39659.0, 39660.0, 39666.0, 39668.0, 39669.0, 39670.0, 39676.0,
E          39678.0, 39679.0],
E         dtype='float64', length=39489)
E   At positional index 2488, first diff: 2496.0 != 2494.0

testing.pyx:173: AssertionError

I think there is still a good chance that cross-version and -platform installs are contributing to this within sgeop, but it will be much more difficult to solve than simply bumping up the tolerance in testing...

[martinfleis]

There seems to be some difference in how something gets computed on Apple Silicon.

[jGaboardi]

Yeah Apple Silicon sticks out, but there seem to be differences across all CI environments.

[jGaboardi]

@martinfleis after some blood, sweat, & tears I've set up our action to curate the observed simplified network data as artifacts so at least we can easily compare between OS & package versions for results.

[martinfleis]

Good job! I think that there won't be a ton of what we could do here as this is most likely due to some precision differences in underlying C code (in Qhull, GEOS...) But good to have the ability to debug.

[martinfleis]

I looked into the differences. Vast majority is precision thing. But there are some visible differences here and there. Not sure how much they matter though and if there's any chance we could prevent those.

My proposal is to check the equality of simplified stuff only on one OS in CI.

[jGaboardi]

My proposal is to check the equality of simplified stuff only on one OS in CI.

Sounds good. I am thinking latest & Apple Silicon since that's what I have, but I am open to another version/OS combo.

As for non-equality testing of the full FUA simplification, is there anything else we should/could check for?

[jGaboardi]

note to self: open another issue to include some documentation about this:

... due to some precision differences in underlying C code (in Qhull, GEOS...)
... is precision thing. But there are some visible differences here and there.

[martinfleis]

Sounds good. I am thinking latest & Apple Silicon since that's what I have, but I am open to another version/OS combo.

I'd probably do ubuntu as that is what anyone can get in a container. Apple Silicon is a bit exclusionary.

[jGaboardi]

As for non-equality testing of the full FUA simplification, is there anything else we should/could check for?

@martinfleis Do you have anything opinions/ideas for this? I suppose we a try testing .shape but even that might be different. I will try that here in another commit.

[martinfleis]

Even shape is different in some cases... I don't know what to check. Rough sum of distances?

[jGaboardi]

hmmmm. Let's think more on this.

[jGaboardi]

Rough sum of distances?

I'll push this up in a commit and swap out the current "known" simplified for that produced by Ubuntu latest. See where that gets us.

[jGaboardi]

We've got 5/7 passing envs by testing against the known results from Ubuntu Python 3.12 latest

[jGaboardi]

I want to get the conftest.py started in #27 merged, then build the logic here to only do equality tests from certain CI environments.