center_crop outputs wrong sized array if provided with odd-numbered dimensions smaller than requested crop size

[mattlbeck]

System Info

transformers 4.40.1, python 3.12

Who can help?

@amyeroberts

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

from transformers.image_transforms import center_crop
import numpy as np
small_even = np.ones((3,4,4))
small_odd = np.ones((3,3,3))
big_even = np.ones((3,10,10))
big_odd = np.ones((3,11,11))
print([center_crop(x, (8,8)).shape for x in [small_even, big_even, big_odd, small_odd]])

Result:

[(3, 8, 8), (3, 8, 8), (3, 8, 8), (3, 7, 7)]

Expected behavior

All arrays should be cropped to the requested size of (8,8). However, odd-numbered dimensions that are smaller than the crop size and require padding result in an unexpected off-by-one output size.

[mattlbeck]

I believe the issue is more accurately caused by odd-numbered difference between original size and new size. Rounding up rather than down when calculating the padding fixes the above test cases.