Put load_image function in image_utils.py & fix image rotation issue

[mishig25]

What does this PR do?

1. Put load_image function in image_utils.py
2. Encountered and fixing PIL.Image.open is rotating jpeg images PIL.Image.open is rotating jpeg images python-pillow/Pillow#4703

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[Narsil]

Which version of Pillow are you using ? transformers doesn't specify a version.

Pillow 8.0.0 is automatically rotating images upon .open based on EXIF Orientation tag.
Also probably a good time to remove code duplication here.

[mishig25]

>>> PIL.__version__
'8.3.1'

also, I get rotated mask results from Inference API, which is how I encountered this issue

[Narsil]

I think we need to add some form of test to expose the necessity for this as it's not trivial and might be fixed upstream (as it claims to be).

So we need an image file (can be in hf-internal-testing with said exif data) and show that the loading is improper (this test can be entirely separated from the pipeline, here we're testing the load_image method, so we can have a separate class of test for this that focuses on specifically this feature.

[mishig25]

@Narsil Should I use this opportunity to address (src):

It might be time to create a utils file and put it in there maybe so we can just reuse this code all the time and test it separately. (it should be another PR, just mentioning it here)

If so, should I just make load_image a method of ImageFeatureExtractionMixin & make Vision pipelines inherit ImageFeatureExtractionMixin?

[Narsil]

Just put the function outside the class, it's a simple function it doesn't deserve a mixin.

In another branch I added src/transformers/audio_utils.py to host ffmpeg_read for instance.

(It's the first thing that popped in mind in terms of location, maybe we can find a better location ultimately, here goal is just to reduce repetition)

[mishig25]

@Narsil please feel free to re-review this PR. I've added the changes as suggested.
Specifically, this test case

transformers/tests/test_image_utils.py

Lines 436 to 442 in dd6d891

	def test_load_img_exif_transpose(self):
	import datasets
	dataset = datasets.load_dataset("hf-internal-testing/fixtures_image_utils", "image", split="test")
	img_file = dataset[3]["file"]
	img_without_exif_transpose = PIL.Image.open(img_file)

[NielsRogge]

LGTM! It's indeed better to add the load_image function into image_utils.py.

[sgugger]

Makes sense to move it there!

[LysandreJik]

Great change, LGTM

[LysandreJik]

Before merging, there seems to be a few tests failing linked to your PR!

[mishig25]

@LysandreJik thanks for notifying about the failing tests. Solved! There was an issue with the hf-internal-testing/fixtures_image_utils I've created