`to_pil_image` rounds down

[rb-synth]

🐛 Describe the bug

The current code is: pic = pic.mul(255).byte().

If the input is, say, 0.9999, this will be rounded down to 254 instead of 255.

>>> torch.as_tensor([0.9999]).mul(255).byte().item()
254

I would suggest that the code be: pic = pic.mul(255).round().byte():

>>> torch.as_tensor([0.9999]).mul(255).round().byte().item()
255

Versions

N/A

[philipsgithub]

hello from my side:)

the behavior discussed in the issue is still observable, although the code is now using numpy torchvision/transforms/functional.py#L282.

since I was testing some back and forth conversions for images (uint8 <-> float32), I noticed the Pixel value differences due to numerical fluctuations quite frequently. switching to double, eg. when doing normalizations etc., helped, but I want to avoid that.

I think with an integrated round() the behavior would be much more intuitive. however, in #8142 it was criticized for being unnecessary overhead if slightly alternating behavior is acceptable.
for others it is the other way around: an extra operation is acceptable and the alternating behavior is not.

to allow both groups to work with the same code (maybe an expert can estimate the group sizes), I see one easy solution for now:

• extend the docstring to mention that behavior so that the users become more aware (maybe also hint to torchvision/transforms/_functional_tensor.py#L90)

but maybe @fmassa or @pmeier have an opinion on that? since the disscussion here #2078 went in a similar direction

[philipsgithub]

my minimal example:

import numpy as np
img_original = np.random.randint(0, 256, (224, 224, 3), dtype=np.uint8)
img_transformed = (img_original/(255 + 1e-13)) # let's say we did some processing on the float image causing small positive numerical deviations
print(np.max(img_transformed), np.min(img_transformed)) # 0.9999999999999996 0.0
img_data_for_pil = (img_transformed * 255).astype(np.uint8) # equivalent to line 282
avg_diff = np.sum(np.abs(img_data_for_pil*1. - img_original)) / img_original.size
print(avg_diff) # >0.99

[YeonwooSung]

I'll work on this: round when converting float images to uint8 in to_pil_image so values like 0.9999 map to 255.

[YeonwooSung]

Opened a fix in #9646 — to_pil_image now rounds float values to nearest uint8 (with clip) so 0.9999 maps to 255.