Change default of antialias parameter from None to 'warn' #7160

NicolasHug · 2023-02-01T17:48:24Z

Before changing the default to True in 2 versions

NicolasHug

@pmeier @vfdev-5 this is still WIP but I'm curious about your initial thoughts on the points noted above

NicolasHug · 2023-02-01T17:48:50Z

torchvision/transforms/functional.py

+# TODO in v0.17: remove this helper and change default of antialias to True everywhere
+def _check_antialias(antialias: Optional[Union[str, bool]], interpolation: InterpolationMode) -> Optional[bool]:
+    if isinstance(antialias, str):  # it should be "warn", but we don't bother checking against that
+        # TODO Should we only warn if input is a tensor, or for PIL images too?


I would only warn for tensors. Since the value is, and will be, ignored for PIL and we already warn that it is passed in the first place

vision/torchvision/transforms/functional.py

Lines 474 to 476 in 135a0f9

if not isinstance(img, torch.Tensor):

if antialias is not None and not antialias:

warnings.warn("Anti-alias option is always applied for PIL Image input. Argument antialias is ignored.")

the warning here is irrelevant for the user.

OK, we're only warning for Tensors now. One minor caveat is that since we can't warn in __init__() of Resize() anymore, the warning will only happen during forward(). I assume that's fine? Right now, I tend to prefer that to the alternative of warning for PIL users, who shouldn't have to be bothered about all this

NicolasHug · 2023-02-01T17:54:56Z

torchvision/transforms/functional_tensor.py

+        # We manually set it to False to avoid an error downstream in interpolate()
+        # This behaviour is documented: the parameter is irrelevant for modes
+        # that are not bilinear or bicubic. We used to raise an error here, but
+        # now we don't as True is the default.


We used to raise an error, now we silently disable antialias. Otherwise, we'd be breaking those use-cases when we switch the default to True in 2 versions:

Resize(interpolation=NEAREST)(img) # uh oh, default changed to True, which raises an error!

This is a slight change of behaviour (not a change of results though!) and it's not perfect, but I feel like the alternative is worse: we could introduce another value for antialias e.g. "auto", which would mean "Do what PIL does i.e. apply antialias for bilinear and bicubic, and no antialias otherwise". But having that + True + False + None will bloat the API IMHO.

If it makes it any better, remember that PIL already silently switches antialiasing on or off depending on the interpolation mode (granted, they don't have an antialias parameter so it's not a completely fair comparison).

torchvision/prototype/datapoints/_bounding_box.py

NicolasHug

I still need to add tests on the prototype transforms, but this is ready for a first round.

Unsurprisingly I found it to be a fairly difficult change. @pmeier @vfdev-5 please don't assume I did the right thing, I need a careful review on this :)

NicolasHug · 2023-02-02T15:24:08Z

test/test_models.py

+        models.detection.FasterRCNN_ResNet50_FPN_Weights.DEFAULT.transforms()(img)
+        models.video.R3D_18_Weights.DEFAULT.transforms()(img)
+        models.optical_flow.Raft_Small_Weights.DEFAULT.transforms()(img, img)


For these 3, AFAICT, Resize() is either not used in the evaluation presets, or the training was done on Tensors already (i.e. no interpolation was done anyway, in which case I hard-coded antialias=False and added comments in the code to explain why).

test/test_functional_tensor.py

torchvision/transforms/transforms.py

NicolasHug · 2023-02-02T18:21:14Z

test/test_prototype_transforms.py

+
+
+# TODO: remove this test in 0.17 when the default of antialias changes to True
+def test_antialias_warning():


I'm mixing the functional and class tests here... Hope that's fine (I found it to be a awkward to scatter them across so many files as done in the V1 tests)

Fine by me. Test is self-contained enough that there shouldn't be any confusion about it later.

…s_default

NicolasHug · 2023-02-06T15:48:58Z

I just confirmed on the references that the behaviour looks fine. I'm getting warnings when relevant (only on tensors) and the accuracy boost is consistent with #6506 (I only checked mobilenet_v3_large this time). On the segmentation references, there's no perf difference on the model I tried (FCN resnet 50). On the other references like segmentation or video, the resizing was done directly on tensors (in the models or in the presets), so there's no change there.

torchvision/prototype/datapoints/_bounding_box.py

torchvision/transforms/functional.py

pmeier · 2023-02-06T19:19:14Z

torchvision/transforms/functional.py

    if not isinstance(img, torch.Tensor):
        if antialias is not None and not antialias:
            warnings.warn("Anti-alias option is always applied for PIL Image input. Argument antialias is ignored.")


Now that we have four possible values, should we refactor these checks? They are quite hard to read TBH:

antialias a = antialias is not None b = not antialias a and b

None False True False

False True True True

True True False False

"warm" True False False

Unless I'm missing something, that should be equivalent to antialias is False?

Yes, is False is clearer, althought I avoided doing this in this PR because it's not strictly needed for the original purpose, and the PR is already pretty long. I don't mind doing if you think it's less confusing

I don't care if we do it here or in a follow-up, but we should do it. I had to build the table above to see if the new case of "warn" does something unexpected and that is not a good sign for an otherwise simple check.

@pmeier antialias="warn" should behave as antialias=None, so, I would say there is no 4th line in the table (also values of the 4th line seem to be incorrect).

Well, the result is the same, but the way it is achieved is different. What do you think is wrong?

>>> antialias = "warn" >>> antialias is not None True >>> not antialias False >>> antialias is not None and not antialias False

…s_default

test/test_functional_tensor.py

vfdev-5 · 2023-02-13T17:17:56Z

test/test_functional_tensor.py

-    resized_tensor = F.resize(tensor, size=size, interpolation=interpolation, max_size=max_size)
-    resized_pil_img = F.resize(pil_img, size=size, interpolation=interpolation, max_size=max_size)
+    resized_tensor = F.resize(tensor, size=size, interpolation=interpolation, max_size=max_size, antialias=True)
+    resized_pil_img = F.resize(pil_img, size=size, interpolation=interpolation, max_size=max_size, antialias=True)


Nit: As we now doing antialias=True, we can also reduce the tol below now:

_assert_approx_equal_tensor_to_pil(resized_tensor_f, resized_pil_img, tol=8.0)

OK, I'll merge now to unblock, and I'll open a PR to check the CI on a lower tol
EDIT: #7233

vfdev-5

OK to me, thanks @NicolasHug
I left few nit picking comments but they can be also ignored.

…s_default

…7160) Summary: Reviewed By: vmoens Differential Revision: D44416275 fbshipit-source-id: 916691a68545d0b487d9ac20b4a8f42ec42315b6 Co-authored-by: Philip Meier <github.pmeier@posteo.de> Co-authored-by: vfdev <vfdev.5@gmail.com>

Change default or antialias parameter from None to 'warn'

71283e8

facebook-github-bot added the cla signed label Feb 1, 2023

NicolasHug commented Feb 1, 2023

View reviewed changes

NicolasHug added 6 commits February 2, 2023 11:20

Don't warn for PIL images

d21e2c2

Remove TODO

ca7932a

Handle presets

e5f619e

Update V2 defaults

4e3624d

Fix some tests

3eb3e30

some stuff not needed anymore

c1d39e7

NicolasHug marked this pull request as ready for review February 2, 2023 15:20

NicolasHug requested review from pmeier and vfdev-5 February 2, 2023 15:20

vfdev-5 reviewed Feb 2, 2023

View reviewed changes

torchvision/prototype/datapoints/_bounding_box.py Show resolved Hide resolved

NicolasHug commented Feb 2, 2023

View reviewed changes

Phew - add test for v2 transforms

daacc46

NicolasHug commented Feb 2, 2023

View reviewed changes

NicolasHug changed the title ~~Change default or antialias parameter from None to 'warn'~~ Change default of antialias parameter from None to 'warn' Feb 6, 2023

NicolasHug added 2 commits February 6, 2023 15:38

Merge branch 'main' of github.com:pytorch/vision into change_antialia…

911db0c

…s_default

Error message

b4921e3

pmeier reviewed Feb 6, 2023

View reviewed changes

pmeier and others added 3 commits February 10, 2023 12:30

get rid of warnings inside test_prototype_tranforms

f6c8140

Merge branch 'main' of github.com:pytorch/vision into change_antialia…

22fd450

…s_default

None -> True in tests

09226ac

NicolasHug mentioned this pull request Feb 10, 2023

TODOs before 0.15 release #7217

Closed

49 tasks

NicolasHug and others added 2 commits February 10, 2023 15:54

Merge branch 'main' into change_antialias_default

8e5e8b6

Merge branch 'main' into change_antialias_default

fc6375b

vfdev-5 reviewed Feb 13, 2023

View reviewed changes

test/test_functional_tensor.py Outdated Show resolved Hide resolved

vfdev-5 reviewed Feb 13, 2023

View reviewed changes

vfdev-5 approved these changes Feb 13, 2023

View reviewed changes

NicolasHug added 3 commits February 13, 2023 17:36

Merge branch 'main' of github.com:pytorch/vision into change_antialia…

5a799fa

…s_default

Remove antialias=True as per suggestion

60c4165

lint

058c6b5

NicolasHug merged commit b030e93 into pytorch:main Feb 13, 2023

NicolasHug added module: transforms deprecation and removed cla signed labels Feb 13, 2023

This was referenced Feb 13, 2023

Tighter test against PIL on resize #7233

Merged

Use is False for some antialias checks #7234

Merged

boegel mentioned this pull request Apr 18, 2024

{ai}[foss/2023a] PyTorch-bundle v2.1.2 easybuilders/easybuild-easyconfigs#20382

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Change default of antialias parameter from None to 'warn' #7160

Change default of antialias parameter from None to 'warn' #7160

NicolasHug commented Feb 1, 2023 •

edited by pytorch-bot bot

Loading

NicolasHug left a comment

NicolasHug Feb 1, 2023

pmeier Feb 2, 2023

NicolasHug Feb 2, 2023 •

edited

Loading

NicolasHug Feb 1, 2023 •

edited

Loading

NicolasHug left a comment

NicolasHug Feb 2, 2023

NicolasHug Feb 2, 2023 •

edited

Loading

pmeier Feb 6, 2023

NicolasHug commented Feb 6, 2023

pmeier Feb 6, 2023 •

edited

Loading

NicolasHug Feb 7, 2023

pmeier Feb 7, 2023

vfdev-5 Feb 13, 2023

pmeier Feb 13, 2023

vfdev-5 Feb 13, 2023

NicolasHug Feb 13, 2023 •

edited

Loading

vfdev-5 left a comment

	if not isinstance(img, torch.Tensor):
	if antialias is not None and not antialias:
	warnings.warn("Anti-alias option is always applied for PIL Image input. Argument antialias is ignored.")



		# TODO: remove this test in 0.17 when the default of antialias changes to True
		def test_antialias_warning():

`antialias`	`a = antialias is not None`	`b = not antialias`	`a and b`
`None`	`False`	`True`	`False`
`False`	`True`	`True`	`True`
`True`	`True`	`False`	`False`
`"warm"`	`True`	`False`	`False`

Change default of antialias parameter from None to 'warn' #7160

Change default of antialias parameter from None to 'warn' #7160

Conversation

NicolasHug commented Feb 1, 2023 • edited by pytorch-bot bot Loading

NicolasHug left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

NicolasHug Feb 2, 2023 • edited Loading

Choose a reason for hiding this comment

NicolasHug Feb 1, 2023 • edited Loading

Choose a reason for hiding this comment

NicolasHug left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

NicolasHug Feb 2, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

NicolasHug commented Feb 6, 2023

pmeier Feb 6, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

NicolasHug Feb 13, 2023 • edited Loading

Choose a reason for hiding this comment

vfdev-5 left a comment

Choose a reason for hiding this comment

NicolasHug commented Feb 1, 2023 •

edited by pytorch-bot bot

Loading

NicolasHug Feb 2, 2023 •

edited

Loading

NicolasHug Feb 1, 2023 •

edited

Loading

NicolasHug Feb 2, 2023 •

edited

Loading

pmeier Feb 6, 2023 •

edited

Loading

NicolasHug Feb 13, 2023 •

edited

Loading