Re-enable or remove test_cpp_models.py?

[NicolasHug]

Currently, all the tests in test_cpp_models.py are unconditionally skipped:

@unittest.skipIf(
    sys.platform == "darwin" or True,
    "C++ models are broken on OS X at the moment, and there's a BC breakage on master; "
    "see https://github.com/pytorch/vision/issues/1191")

(note the or True).

I tried running them on my laptop and they all passed except for a few of the mnasnet tests that failed with RuntimeError: Given groups=32, expected weight to be divisible by 32 at dimension 0, but got weight of size [[40, 1, 3, 3]] instead , which is probably fairly easy to fix. BTW, I'm on OS X so the sys.platform == "darwin" can probably be removed too.

Should we fix and re-activate these tests? Or perhaps they are dedundant with the ones in test_models.py now? The tests ran in 2+ minutes for me, so reactivating them will have a significant impact on CI usage, which is not negligible.

CC @fmassa

cc @pmeier

[fmassa]

Those tests test the C++ API of the models in torchvision. The current way we did to test them is not ideal, as we expose the C++ models to Python to compare the results of the baseline implementation in Python.

In principle we should enable the C++ tests back, but given that the C++ API of torchvision is in a state which it is not maintained (we don't enforce adding C++-equivalent models to the new Python models added), I'm tempted to not spend too much effort with it. If it's a quick fix, then great, but let's make sure we don't ship our torchvision binaries with the C++ models, and that it is only compiled in the unittests.

Additionally, it would be good to know who is the current maintainer of the C++ API of PyTorch

[NicolasHug]

@fmassa , now that #4375 is merged and that the cpp API is deprecated, should we just remove the tests altogether? They're always skipped at the moment (and have been for a while)

[NicolasHug]

I don't see this file anymore, we probably removed it