CPU unit tests are failing in vision nightly builds

[izaitsevfb]

🐛 Describe the bug

HUD: https://hud.pytorch.org/hud/pytorch/vision/nightly/1?per_page=50&name_filter=tests%20%2F%20unittests

Looks like the same issue across platforms:

2023-10-16T11:57:47.9045908Z �[31mFAILED�[0m test/test_utils.py::�[1mtest_draw_boxes�[0m - AssertionError: Tensor-likes are not equal!
2023-10-16T11:57:47.9046508Z 
2023-10-16T11:57:47.9046642Z Mismatched elements: 240 / 30000 (0.8%)
2023-10-16T11:57:47.9047134Z Greatest absolute difference: 251 at index (0, 10, 5)
2023-10-16T11:57:47.9047689Z Greatest relative difference: inf at index (0, 8, 4)
2023-10-16T11:57:47.9048831Z �[31m= �[31m�[1m1 failed�[0m, �[32m21402 passed�[0m, �[33m16844 skipped�[0m, �[33m1 xfailed�[0m, �[33m251 warnings�[0m�[31m in 1206.38s (0:20:06)�[0m�[31m =�[0m
2023-10-16T11:57:47.9062170Z ##[error]Process completed with exit code 1.

cc @NicolasHug @pmeier @atalman

Versions

null

[pmeier]

We haven't merged anything the last few days. Meaning, this likely comes from core. Let me dig and see if I can find an offending PR.

[pmeier]

Could also come from Pillow. Error points to

vision/test/test_utils.py

Lines 108 to 111 in 1d6f787

	if PILLOW_VERSION >= (8, 2):
	# The reference image is only valid for new PIL versions
	expected = torch.as_tensor(np.array(Image.open(path))).permute(2, 0, 1)
	assert_equal(result, expected)

and Pillow==10.1 was released two days ago.

[pmeier]

Can reproduce with Pillow==10.1.0.

[pmeier]

The default font has changed. See python-pillow/Pillow#7354

reference / new

[pmeier]

There currently seems to be no way to force the old font. I'll ask for clarification there. In the mean time, we could temporarily pin to pillow<10.1 until we have a plan.