Use stable sort in nms

[shuokay]

🚀 The feature

Please use stable sort in nms since torch1.9 has supported stable sort.

Motivation, pitch

Why we need stable sort:
In quantization model, many score are equal due to quantization, so when using unstable sort, the result of nms is non-reproducible.

Alternatives

No response

Additional context

No response

[datumbox]

It's true that the sort component of NMS has been the source of many headaches. The fact that the algorithm is unstable means that we get different order depending on platform/python/system config. This leads to us being unable to properly test those algorithms and we often resort to ugly hacks/workarounds:

vision/test/test_models.py

Lines 579 to 592 in 2e0949e

	try:
	# We first try to assert the entire output if possible. This is not
	# only the best way to assert results but also handles the cases
	# where we need to create a new expected result.
	_assert_expected(output, model_name, prec=prec)
	except AssertionError:
	# Unfortunately detection models are flaky due to the unstable sort
	# in NMS. If matching across all outputs fails, use the same approach
	# as in NMSTester.test_nms_cuda to see if this is caused by duplicate
	# scores.
	expected_file = _get_expected_file(model_name)
	expected = torch.load(expected_file)
	torch.testing.assert_close(output[0]["scores"], expected[0]["scores"], rtol=prec, atol=prec,
	check_device=False, check_dtype=False)