Make "RuntimeError: operator torchvision::nms does not exist" error better

[zou3519]

I'm putting this issue in PyTorch repo because I'm not sure where the fix actually goes.
If some shared library is missing, and torchvision._C fails to import, then I get the above error.
Can we do something about this to make the error message clearer?

cc @atalman @NicolasHug

[malfet]

IMO fix should go into torchvision, which since times immemorial had this operational mode when it's partially usable without _C library (probably to support loading resnet thru torchhub), but not sure if this is still relevant

[NicolasHug]

@zou3519 would you have a reproducing example by any chance? I'm trying to understand why we don't hit this better error message first:

vision/torchvision/extension.py

Lines 44 to 54 in 1e53952

	def _assert_has_ops():
	if not _has_ops():
	raise RuntimeError(
	"Couldn't load custom C++ ops. This can happen if your PyTorch and "
	"torchvision versions are incompatible, or if you had errors while compiling "
	"torchvision from source. For further information on the compatible versions, check "
	"https://github.com/pytorch/vision#installation for the compatibility matrix. "
	"Please check your PyTorch version with torch.__version__ and your torchvision "
	"version with torchvision.__version__ and verify if they are compatible, and if not "
	"please reinstall torchvision so that it matches your PyTorch install."
	)

since it should be called before accessing nms:

vision/torchvision/ops/boxes.py

Lines 47 to 48 in 1e53952

	_assert_has_ops()
	return torch.ops.torchvision.nms(boxes, scores, iou_threshold)

---

Regardless, I'd love to raise loudly instead of silencing errors when loading the ops in

vision/torchvision/extension.py

Lines 40 to 41 in 1e53952

	except (ImportError, OSError):
	pass

I think you're right @malfet that this was meant for torchhub, but it might be time to break that.