Failing tests in tests suite

[VRehnberg]

I'm working on an easybuild recipe for installing PyTorch3D from source and encountered a handful of bugs with the test-suite in the process. I don't have time to file proper issues right now, but figured I can at least share them as is.

Bugs with patches

1. tests/implicitron/test_config.py Bar dataclass immutable patch
2. tests/implicitron/test_extending_orm_types.py load_blobs not positional in build method patch
3. tests/pulsar/test_forward.py issue with 1 channel images patch

Bugs without patches

1. tests/test_io_off.py::TestMeshOffIO::test_bad and tests/test_io_off.py::TestMeshOffIO::test_load_lumpy, error parsing file
2. tests/test_knn.py::TestKNN::test_knn_vs_python_square_cuda, tests/test_points_alignment.py::TestICP::test_heterogeneous_inputs and tests/test_render_meshes.py::TestRenderMeshes::test_simple_sphere_elevated_camera seems flaky, probably depending on hardware used.

[VRehnberg]

If I remember correctly, many tests also assumed there were multiple GPUs available, which will fail if building in single GPU jobs.

PyTorch test-suite handles this with a pytest.SkipIf https://github.com/pytorch/pytorch/blob/main/torch/testing/_internal/common_cuda.py#L17

[bottler]

The OFF problems should have been fixed yesterday by dd2a11b . The code is unfortunately dependent on the wording of numpy's exceptions which have changed. The conditional use of loadtxt seems to me the most efficient implementation of loading off files which are likely to be homogeneous.

[bottler]

This is very useful! I need to learn what easybuild is! Yes, there are some flaky tests as you mention, which we've just been leaving failing.

[bottler]

The simple fixes you suggest have now landed. Thank you for them. How much of a blocker are the other things?

[VRehnberg]

Not a blocker, the related tests are skipped manually. Only impact is less test coverage to catch any actual errors with a build.