[TST] Upcoming dependency test failures #27892

Activity

1. ksunden commented on Mar 12, 2024

   

   ksunden

   on Mar 12, 2024

   Member

   More actions

   This test was run during a particularly high churn portion of the numpy 2 release cycle, so I'm guessing most of these failures are related to that.

   Numpy has released 2.0.0b1 now, and testing locally I see a couple of these, but not nearly as many as the CI job found:

   FAILED lib/matplotlib/tests/test_triangulation.py::test_trifinder - AssertionError: 
   FAILED lib/matplotlib/tests/test_triangulation.py::test_triinterpcubic_C1_continuity - AssertionError: 
   FAILED lib/matplotlib/tests/test_triangulation.py::test_triinterp - AssertionError: 
   FAILED lib/matplotlib/tests/test_contour.py::test_all_algorithms[png-False] - matplotlib.testing.exceptions.ImageComparisonFailure: images not close (RMS 70.183):
   FAILED lib/matplotlib/tests/test_contour.py::test_all_algorithms[png-True] - matplotlib.testing.exceptions.ImageComparisonFailure: images not close (RMS 70.183):
   FAILED lib/matplotlib/tests/test_triangulation.py::test_trirefine - AssertionError: 
   FAILED lib/matplotlib/tests/test_backends_interactive.py::test_qt_missing - subprocess.CalledProcessError: Command '['/home/kyle/.pyenv/versions/3.11.0/envs/edge/bin/python3.11', '-c', "import importlib.util;_spec = importlib.util.spec_from_file_location('matplotlib.tests.test_backends_interactive', '/home/...
   FAILED lib/matplotlib/tests/test_triangulation.py::test_tri_smooth_gradient[png] - matplotlib.testing.exceptions.ImageComparisonFailure: images not close (RMS 42.063):
   ERROR lib/matplotlib/tests/test_cbook.py::test_reshape2d_xarray - AttributeError: `np.unicode_` was removed in the NumPy 2.0 release. Use `np.str_` instead.
   ERROR lib/matplotlib/tests/test_cbook.py::test_index_of_xarray - AttributeError: `np.unicode_` was removed in the NumPy 2.0 release. Use `np.str_` instead.

   Digging into what I can from here.

   The xarray tests appear to be from not reinstalling pyarrow compiled against np2.0, so not too worried about those. The qt one I think is just a my machine problem that I'll need to sort out at some point.

   I'm most interested in the triangulation problems, which manifest in several tests failed at assertions and two image test failures (counting test_all_algorithms as one).

   The image failures are quite real, looks like two contour algorithms fail to plot at all, and the "test_smooth_gradient" results in arrows that are completely different.
2. ksunden commented on Mar 12, 2024

   

   ksunden

   on Mar 12, 2024

   Member

   More actions

   Okay, I think that these are likely due to pybind/pybind11#5050 (some of which via contourpy/contourpy#369)

   A tad bit surprised that if there are problems with pybind11 we don't see them a little more widespread (path in particular), but we don't.

   In any case, the solutions appear to be well on their way with little to do here directly. (Though it may put a slight pause on us releasing our rc to go with this np beta).

   I will leave this open for now as those get worked through.
3. ianthomas23 commented on Mar 12, 2024

   

   ianthomas23

   on Mar 12, 2024

   Member

   More actions

   Those failures look consistent with what I'm observing in contourpy which is that uint32 arrays are affected but float64 and uint8 don't seem to be.

   If pybind/pybind11#5050 remains as it is you will need to add extra #defines before #include <pybind11/numpy.h> and build against a pybind11 release that includes that PR.
4. ksunden commented on Mar 12, 2024

   

   ksunden

   on Mar 12, 2024

   Member

   More actions

   Looks like there is at least conversation leaning towards ON by default, which would mean no change is needed, but it hasn't made it to the implementation yet from what I can tell...

   👍React with 👍1Reacted by Ian Thomas
5. ianthomas23 commented on Mar 28, 2024

   

   ianthomas23

   on Mar 28, 2024

   Member

   More actions

   FYI ContourPy 1.2.1rc1 is available on PyPI (https://pypi.org/project/contourpy/1.2.1rc1/#files), built using pybind11 2.12.0 and supporting NumPy 2. There will be a full release of ContourPy 1.2.1 after NumPy 2.0.0rc1.

   👍React with 👍1Reacted by Kyle Sunden
6. ianthomas23 commented on Apr 2, 2024

   

   ianthomas23

   on Apr 2, 2024

   Member

   More actions

   Full ContourPy 1.2.1 release is now on PyPI: https://pypi.org/project/contourpy/1.2.1/#files.

   🎉React with 🎉2Reacted by Kyle Sunden and Clément Robert
7. ksunden commented on May 8, 2024

   

   ksunden

   on May 8, 2024

   Member

   More actions

   Sorted out between numpy 2 support/contourpy/pybind11 updates
8. ksunden

   closed this as completedon May 8, 2024