pytorch

3766c047 - Add uint8 support for CPU images in interpolate(mode='bicubic') (#103252)

Add uint8 support for CPU images in interpolate(mode='bicubic') (#103252) CC @vfdev-5 Proposed strategy: Be as close as possible to PIL when `antialias=True`. Be as close as possible to float path when `antialias=False`. Ad-hoc tests: <details> ```py import random import torch import pytest import numpy as np from PIL import Image from torch.nn.functional import interpolate @pytest.mark.parametrize("C", (1, 3, 6)) @pytest.mark.parametrize("batch_size", (1, 4)) @pytest.mark.parametrize("memory_format", (torch.contiguous_format, torch.channels_last, "strided", "cropped")) @pytest.mark.parametrize("antialias", (True, False)) # @pytest.mark.parametrize("mode", ("bilinear", "bicubic",)) @pytest.mark.parametrize("mode", ("bicubic",)) @pytest.mark.parametrize("seed", range(100)) def test_resize(C, batch_size, memory_format, antialias, mode, seed): def test_resize(C, batch_size, memory_format, antialias, mode, 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) img = torch.randint(0, 256, size=(batch_size, C, Hi, Wi), dtype=torch.uint8) if memory_format in (torch.contiguous_format, torch.channels_last): img = img.to(memory_format=memory_format, copy=True) elif memory_format == "strided": 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) img = img[:, :, a:b, c:d] else: raise ValueError("Uh?") margin = 0 img = img.clip(margin, 255 - margin) out_uint8 = interpolate(img, size=[Ho, Wo], mode=mode, antialias=antialias) if antialias and C == 3: out_pil_tensor = resize_with_pil(img, Wo, Ho, mode=mode, antialias=antialias) atol = {"bicubic": 2, "bilinear": 1}[mode] # TODO: is 2 expected when comparing with PIL bicubic? Why not 1 as for bilinear? torch.testing.assert_close(out_uint8, out_pil_tensor, rtol=0, atol=atol) out_float = interpolate(img.to(torch.float), size=[Ho, Wo], mode=mode, antialias=antialias).round().clip(0, 255).to(torch.uint8) if mode == "bicubic": diff = (out_float.float() - out_uint8.float()).abs() assert diff.max() < 30 percent = .03 if antialias else .1 assert (diff > 2).float().mean() < percent mae = .4 if antialias else .8 assert diff.mean() < mae else: torch.testing.assert_close(out_uint8, out_float, rtol=0, atol=1) def resize_with_pil(batch, Wo, Ho, mode, antialias): resample = {"bicubic": Image.BICUBIC, "bilinear": Image.BILINEAR}[mode] out_pil = [ Image.fromarray(img.permute((1, 2, 0)).numpy()).resize((Wo, Ho), resample=resample) for img in batch ] out_pil_tensor = torch.cat( [ torch.as_tensor(np.array(img, copy=True)).permute((2, 0, 1))[None] for img in out_pil ] ) return out_pil_tensor ``` </details> Pull Request resolved: https://github.com/pytorch/pytorch/pull/103252 Approved by: https://github.com/vfdev-5, https://github.com/H-Huang, https://github.com/malfet, https://github.com/atalman