Fix bugs in mllama image processing #36156
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: Travis Johnson <[email protected]>
Signed-off-by: Travis Johnson <[email protected]>
Member
|
gentle ping @qubvel |
qubvel
reviewed
Mar 6, 2025
Contributor
qubvel
left a comment
qubvel
left a comment
There was a problem hiding this comment.
Hi @tjohnson31415, thanks for the PR, I left a few comments
Comment on lines
243
to
247
| image_inputs = [[Image.new("RGB", (100, 1))]] | ||
| encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values | ||
| expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs) | ||
| self.assertEqual(tuple(encoded_images.shape), (1, *expected_output_image_shape)) | ||
|
|
Contributor
There was a problem hiding this comment.
can you also add a test with 1x1 image for PIL?
Co-authored-by: Pavel Iakubovskii <[email protected]>
Signed-off-by: Travis Johnson <[email protected]>
Signed-off-by: Travis Johnson <[email protected]>
Signed-off-by: Travis Johnson <[email protected]>
tjohnson31415
force-pushed
the
fix-mllama-image-processing
branch
from
c162ced to
d22cd02
Compare
Contributor
Author
|
Thanks @Rocketknight1 for the ping and @qubvel for the review! I have made the suggested changes. I also added a fix for another edge-case bug with an image with an impractical aspect ratio that would result in the resized image having an invalid dimension of 0. Please review this additional fix and test as well 🙏 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
I found that using a 1x1 PIL image with the
MllamaImageProcessorsurfaced a couple of bugs:A 1x1 PIL image results in ambgiuous channel dimension in
infer_channel_dimension_format. The default in the ambiguous case is to use the first dimension, which is incorrect and results in aValueError: mean must have 1 elements if it is an iterable, got 3. Another bug was that explicility setting the input channel format results in a similar error:ValueError: mean must have 224 elements if it is an iterable, got 3.This PR resolves the bugs by:
input_data_format="channels_last"by usingdata_formatinstead ofinput_data_formatfor rescaling and normalizing after the call toto_channel_dimension_formatThere is some discussion in an issue reporting the same bug that was closed without a fix: #34029
Another edge case bug this PR fixes is handling an image that has an impractical aspect ratio instead of having PIL raise a
ValueError: height and width must be > 0error. The fix here is to force the minimum resized image dimensions to be 1 or greater.Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
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.
@ArthurZucker
@qubvel