Redundant to_channel_dimension_format() call makes preprocessing fail in case the image has height of 1 pixel#20728
Merged
sgugger merged 1 commit intohuggingface:mainfrom Dec 13, 2022
Conversation
`image = to_channel_dimension_format(image, ChannelDimension.LAST)`
is redundant as this same conversion is also applied in to_pil_image().
This redundant call actually makes the training fail in rare cases.
The problem can be reproduced with the following code snippet:
```
from transformers.models.clip import CLIPFeatureExtractor
vision_processor = CLIPFeatureExtractor.from_pretrained('openai/clip-vit-large-patch14')
images = [
torch.rand(size=(3, 2, 10), dtype=torch.float),
torch.rand(size=(3, 10, 1), dtype=torch.float),
torch.rand(size=(3, 1, 10), dtype=torch.float)
]
for image in images:
processed_image = vision_processor(images=image, return_tensors="pt")['pixel_values']
print(processed_image.shape)
assert processed_image.shape == torch.Size([1, 3, 224, 224])
```
The last image has a height of 1 pixel.
The second call to to_channel_dimesion_format() will transpose the image, and the height
dimension is wrongly treated as the channels dimension afterwards.
Because of this, the following normalize() step will result in an
exception.
|
The documentation is not available anymore as the PR was closed or merged. |
Collaborator
|
cc @amyeroberts |
amyeroberts
approved these changes
Dec 12, 2022
Contributor
amyeroberts
left a comment
There was a problem hiding this comment.
Thanks for finding the issue and fix!
Contributor
Author
|
sure thing! |
mpierrau
pushed a commit
to mpierrau/transformers
that referenced
this pull request
Dec 15, 2022
…ngface#20728) `image = to_channel_dimension_format(image, ChannelDimension.LAST)` is redundant as this same conversion is also applied in to_pil_image(). This redundant call actually makes the training fail in rare cases. The problem can be reproduced with the following code snippet: ``` from transformers.models.clip import CLIPFeatureExtractor vision_processor = CLIPFeatureExtractor.from_pretrained('openai/clip-vit-large-patch14') images = [ torch.rand(size=(3, 2, 10), dtype=torch.float), torch.rand(size=(3, 10, 1), dtype=torch.float), torch.rand(size=(3, 1, 10), dtype=torch.float) ] for image in images: processed_image = vision_processor(images=image, return_tensors="pt")['pixel_values'] print(processed_image.shape) assert processed_image.shape == torch.Size([1, 3, 224, 224]) ``` The last image has a height of 1 pixel. The second call to to_channel_dimesion_format() will transpose the image, and the height dimension is wrongly treated as the channels dimension afterwards. Because of this, the following normalize() step will result in an exception.
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
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.
In the
resize()function in image_transforms.py, the line 267: I thinkimage = to_channel_dimension_format(image, ChannelDimension.LAST)is redundant as this conversion is also applied in the followingto_pil_image().This redundant call actually makes the clip preprocessing fail in special cases. The problem can be reproduced with the following code snippet:
The last image has a height of 1 pixel.
The second call to
to_channel_dimesion_format()will transpose the image, and the height dimension is wrongly treated as the channels dimension afterwards. Because of this, the following normalize() step will result in an exception.An image of height 1 pixel honestly doesn't make much sense, but it happened in my training on visual genome region descriptions and took me a while to track down the problem.
What does this PR do?
Fixes # (issue)
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.