Resize V2 relies on interpolate's native uint8 handling

[vfdev-5]

Description:

• Now that Add uint8 support for interpolate for CPU images pytorch#90771 is merged, let Resize() rely on interpolate()'s native uint8 handling instead of converting to and from float.

  • uint8 input is not casted to f32 for nearest mode (cpu and cuda) and bilinear mode if the latter has AVX2, cpu only.

Context: #7217 and #7497

Benchmarks:

[-- Resize cpu torch.uint8 InterpolationMode.NEAREST -]
                         |  resize v2  |  resize stable
1 threads: --------------------------------------------
      (3, 400, 400)      |      457    |        487    
      (16, 3, 400, 400)  |     6880    |      10100    

Times are in microseconds (us).

[- Resize cpu torch.uint8 InterpolationMode.BILINEAR -]
                         |  resize v2  |  resize stable
1 threads: --------------------------------------------
      (3, 400, 400)      |      330    |        849    
      (16, 3, 400, 400)  |     4420    |      14900    

Times are in microseconds (us).

Source

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/7557

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 2 New Failures

As of commit 0e27ad8:

NEW FAILURES - The following jobs have failed:

• c-source / linux-job (gh)
• pytorch/vision / manywheel-py3_8-rocm5_3 (gh)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[NicolasHug]

Thanks @vfdev-5 , PR looks good. I made a few comments, LMK what you think

[NicolasHug]

Alright, tests are passing and we also stress-tested the changes locally against the float implementation. Looks safe to merge now.

Thanks a ton @vfdev-5 for all your work on interpolate() and with bug chasing / fixing!

import random

import torch
from torchvision.transforms import Resize as Resize_v1
from torchvision.transforms.v2 import Resize as Resize_v2
import pytest

@pytest.mark.parametrize("C", range(1, 6))
@pytest.mark.parametrize("batch_size", ("3d", 1, 4))
@pytest.mark.parametrize("memory_format", (torch.contiguous_format, torch.channels_last, "strided", "cropped"))
@pytest.mark.parametrize("antialias", (True, False))
@pytest.mark.parametrize("seed", range(100))
def test_resize(C, batch_size, memory_format, antialias, seed):
    
    torch.manual_seed(seed)
    random.seed(seed)

    Hi = 2**random.randint(3, 10) + random.randint(0, 30) 
    Wi = 2**random.randint(3, 10) + random.randint(0, 30) 
    Ho = 2**random.randint(3, 10) + random.randint(0, 30) 
    Wo = 2**random.randint(3, 10) + random.randint(0, 30) 
    print(Hi, Wi, Ho, Wo)

    if batch_size == "3d":
        img = torch.randint(0, 256, size=(C, Hi, Wi), dtype=torch.uint8)
    else:
        img = torch.randint(0, 256, size=(batch_size, C, Hi, Wi), dtype=torch.uint8)

    if memory_format in (torch.contiguous_format, torch.channels_last):
        if batch_size == "3d":
            return
        img = img.to(memory_format=memory_format, copy=True)
    elif memory_format == "strided":
        if batch_size == "3d":
            img = img[:, ::2, ::2]
        else:
            img = img[:, :, ::2, ::2]
    elif memory_format == "cropped":
        a = random.randint(1, Hi // 2)
        b = random.randint(Hi // 2 + 1, Hi)
        c = random.randint(1, Wi // 2)
        d = random.randint(Wi // 2 + 1, Wi)
        if batch_size == "3d":
            img = img[:, a:b, c:d]
        else:
            img = img[:, :, a:b, c:d]
    else:
        raise ValueError("Uh?")
    

    out_uint8 = Resize_v2(size=(Ho, Wo), antialias=antialias)(img)
    out_float = Resize_v1(size=(Ho, Wo), antialias=antialias)(img.float()).round().to(torch.uint8)

    torch.testing.assert_close(out_uint8, out_float, rtol=0, atol=1)

        

[github-actions]

Hey @NicolasHug!

You merged this PR, but no labels were added. The list of valid labels is available at https://github.com/pytorch/vision/blob/main/.github/process_commit.py