fix the get_size_with_aspect_ratio in max_size situation

[SangbumChoi]

What does this PR do?

This fix the issue as the following example. AKA round issue.
This is important when the size should be divisible by specific value such as 32, 16.
@amyeroberts

from transformers import DetrImageProcessor
from PIL import Image
import requests
import torch
import io
import numpy as np


def rgb_to_id(color):
    if isinstance(color, np.ndarray) and len(color.shape) == 3:
        if color.dtype == np.uint8:
            color = color.astype(np.int32)
        return color[:, :, 0] + 256 * color[:, :, 1] + 256 * 256 * color[:, :, 2]
    return int(color[0] + 256 * color[1] + 256 * 256 * color[2])
    
height, width = 653, 958
image = Image.new('RGB', (width, height), 'white')
image_processor = DetrImageProcessor.from_pretrained('facebook/detr-resnet-50')
image_processor.size = \
{
    "longest_edge": 640,
    "shortest_edge": 640
}

Before commit

# prepare inputs for the model
inputs = image_processor(images=image, return_tensors="pt")
print(inputs['pixel_values'].shape)
torch.Size([1, 3, 436, 639])

After commit

# prepare inputs for the model
inputs = image_processor(images=image, return_tensors="pt")
print(inputs['pixel_values'].shape)
torch.Size([1, 3, 436, 640])

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.

[SangbumChoi]

Additional script

height, width = 653, 958

for i in range(640,660):
    image = Image.new('RGB', (width, i), 'white')
    
    print(image.size)
    
    image_processor = DetrImageProcessor.from_pretrained('facebook/detr-resnet-50')
    image_processor.size = \
    {
        "longest_edge": 640,
        "shortest_edge": 640
    }
    
    # prepare inputs for the model
    inputs = image_processor(images=image, return_tensors="pt")
    print(inputs['pixel_values'].shape)

(958, 640)
torch.Size([1, 3, 428, 640])
...
(958, 659)
torch.Size([1, 3, 440, 640])

[SangbumChoi]

transformers/src/transformers/models/yolos/image_processing_yolos.py

Line 104 in fce78fd

def get_size_with_aspect_ratio(image_size, size, max_size=None) -> Tuple[int, int]:

We can consider merging both all-in-one

[amyeroberts]

Thanks for fixing!

Could you add a test, checking both the upper and lower bounds, which fail on main and pass here?

cc @qubvel

[qubvel]

@SangbumChoi Thanks for fixing, a great catch!

We can consider merging both all-in-one

will be great to have them consistent, but take into consideration that yolos function "round" height and width to the lowest divisible value

[SangbumChoi]

I will update the yolos and the test function and ping again.

[SangbumChoi]

@qubvel @amyeroberts I added the screenshot which fails at the main branch and success in here.

For the record, I tried to make yolos consistent to all the other function. However, it turns out it needs to remove #Copied from line which I would like to maintain. As a result I just changed it with appropriate format.

I just added test on the detr but if it seems good I could add it to all the related models.

[amyeroberts]

Thanks for working on this! Overall looks good to me. It would be good to add some tests to make sure the behaviour is as expected for specific cases eg. when max_size is set or not; height < width; width < height; height == width; height == size etc.

[SangbumChoi]

I have added the 3 more cases which is total 5 cases. Note that max_size should be always set in the situation of get_size_with_aspect_ratio! Again I just added test on the detr but if it seems good I could add it to all the related models :)

[SangbumChoi]

@amyeroberts Hi amy, since recent merge has deprecated DETA. I also removed in here.

[SangbumChoi]

@amyeroberts Could you rerun the CI?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[amyeroberts]

Thanks for fixing!