torchvision 0.4.2: test_adjust_saturation fails with pillow 7.0.0

[hartb]

Test test_adjust_saturation creates an image from a small array and calls torchvision.transorms.functional.adjust_saturation to adjust the saturation to a couple different levels (0.5 and 2), and then compares the result to some hard-coded expected values.

Torchvision's adjust_saturation routine relies on PIL.ImageEnhance to do the actual work.

The results from pillow's implementation of that routine has changed between pillow 6.2.1 and 7.0.0 (at least in Anaconda's pre-built pillow packages):

$ cat test.py
import numpy as np
from PIL import Image, ImageEnhance, __version__ as PILLOW_VERSION

x_shape = [2, 2, 3]
x_data = [0, 5, 13, 54, 135, 226, 37, 8, 234, 90, 255, 1]

x_np = np.array(x_data, dtype=np.uint8).reshape(x_shape)
x_pil = Image.fromarray(x_np, mode='RGB')

print(f'Pillow version {PILLOW_VERSION}')
for saturation in [0.5, 2]:
    y_pil = ImageEnhance.Color(x_pil).enhance(saturation)
    print(f'adjust_saturation({saturation:3.1f}) => {np.array(y_pil).reshape([12])}')

(pil6) $ python test.py
Pillow version 6.2.1
adjust_saturation(0.5) => [  2   4   8  87 128 173  39  25 138 133 215  88]
adjust_saturation(2.0) => [  0   6  22   0 149 255  32   0 255   4 255   0]

(pil7) $ python test.py
Pillow version 7.0.0
adjust_saturation(0.5) => [  2   4   8  87 128 173  39  25 138 133 216  89]
adjust_saturation(2.0) => [  0   6  22   0 149 255  32   0 255   3 255   0]

Note the differences in the final 3 values between v6 and v7:

• 133 215 88 vs 133 216 89, and
• 4 255 0 vs 3 255 0

The output from the newer pillow version no longer matches the hard-coded values in the test, and so the test fails.

This problem is seen on both x86_64 and ppc64le.

Possible solutions migtht be any of:

• carry canned values for each of pillow v6 and v7
• choose an input with stable outputs across both versions
• relax the allclose assertion to atol=1 to allow for slight differences due to rounding (i.e. to allow 1 least significant bit of difference, since these are uint8 quantities)

[datumbox]

This issue appears to be stale, so I'm closing it. If you think it should remain open, feel free to reopen it.