Fix void segmentation map label reduction

[Yanis01682]

Summary

• preserve existing void/ignore label 255 when
  educe_label is applied
• only reduce non-ignore class ids by 1 for both torch and PIL image processors
• add a SegFormer regression test covering

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: beit, chmv2, dpt, mobilevit, segformer

[Rocketknight1]

cc @molbap @NielsRogge maybe?

[notrudyyy]

Hi @Yanis01682! Could you help me understand what benefits your change to the logic brings?

Unless I am missing something, although the original logic isn't explicit about the fact that it doesn't touch existing 255 labels, it works fine.

If the concern is performance, your proposed method is 2x slower than the original, primarily due to the clone/copy used. If removed, it is still slightly slower. (I could be doing this benchmark wrong, so please correct me if so)

Old code vs new code:

import torch
import timeit

image_base = torch.tensor([0, 1, 255])[torch.randint(0, 3, (32, 32, 3))]
image_base = image_base.to(dtype=torch.uint8)

N = 10_000
images = [image_base.clone() for _ in range(N)]
idx = 0

def method1():
    global idx
    image = images[idx]
    idx += 1
    image[image == 0] = 255
    image -= 1
    image[image == 254] = 255

time1 = timeit.timeit(method1, number=N)

method2_stmt = """
image = image_base.clone()
ignore_mask = (image == 0) | (image == 255)
image[ignore_mask] = 255
image[~ignore_mask] -= 1
"""

time2 = timeit.timeit(
    stmt=method2_stmt,
    globals={"image_base": image_base},
    number=N
)

print(f"Original: {time1:.6f} seconds")
print(f"New: {time2:.6f} seconds")

# Original: 0.672214 seconds
# New: 1.805491 seconds

Old code vs modified new code

import torch
import timeit

image_base = torch.tensor([0, 1, 255])[torch.randint(0, 3, (32, 32, 3))]
image_base = image_base.to(dtype=torch.uint8)

N = 10_000
images = [image_base.clone() for _ in range(N)]
idx = 0

def method1():
    global idx
    image = images[idx]
    idx += 1
    image[image == 0] = 255
    image -= 1
    image[image == 254] = 255

time1 = timeit.timeit(method1, number=N)

images = [image_base.clone() for _ in range(N)]
idx = 0

def method2():
    global idx
    image = images[idx]
    ignore_mask = (image == 0) | (image == 255)
    image[ignore_mask] = 255
    image[~ignore_mask] -= 1

time2 = timeit.timeit(method2, number=N)

print(f"Original: {time1:.6f} seconds")
print(f"New (clone removed): {time2:.6f} seconds")

# Original: 0.705085 seconds
# New (clone removed): 0.777739 seconds

[Yanis01682]

Thanks for taking a close look here. I revisited the original logic and you are right: it already preserves existing 255 labels, so my change does not provide a behavioral improvement there. I also re-checked issue #30064, and the original failure mode was in the Mask2Former void-mask path rather than this SegFormer/related reduce_label logic, and that path appears to have been addressed on main already. Since this PR is not fixing a real remaining issue and adds a slower implementation, I am going to close it. Thanks again for the careful review.