Add uint8 bicubic support to ResizeV2 (#7668)

Co-authored-by: vfdev-5 <vfdev.5@gmail.com>

test/test_transforms_v2_consistency.py

@@ -87,10 +87,8 @@ CONSISTENCY_CONFIGS = [
             ArgsKwargs([32]),
             ArgsKwargs((32, 29)),
             ArgsKwargs((31, 28), interpolation=v2_transforms.InterpolationMode.NEAREST),
-            ArgsKwargs((33, 26), interpolation=v2_transforms.InterpolationMode.BICUBIC),
             ArgsKwargs((30, 27), interpolation=PIL.Image.NEAREST),
             ArgsKwargs((35, 29), interpolation=PIL.Image.BILINEAR),
-            ArgsKwargs((34, 25), interpolation=PIL.Image.BICUBIC),
             NotScriptableArgsKwargs(31, max_size=32),
             ArgsKwargs([31], max_size=32),
             NotScriptableArgsKwargs(30, max_size=100),
@@ -101,6 +99,15 @@ CONSISTENCY_CONFIGS = [
         # atol=1 due to Resize v2 is using native uint8 interpolate path for bilinear and nearest modes
         closeness_kwargs=dict(rtol=0, atol=1),
     ),
+    ConsistencyConfig(
+        v2_transforms.Resize,
+        legacy_transforms.Resize,
+        [
+            ArgsKwargs((33, 26), interpolation=v2_transforms.InterpolationMode.BICUBIC, antialias=True),
+            ArgsKwargs((34, 25), interpolation=PIL.Image.BICUBIC, antialias=True),
+        ],
+        closeness_kwargs=dict(rtol=0, atol=21),
+    ),
     ConsistencyConfig(
         v2_transforms.CenterCrop,
         legacy_transforms.CenterCrop,
@@ -309,15 +316,22 @@ CONSISTENCY_CONFIGS = [
             ArgsKwargs(17, scale=(0.3, 0.7)),
             ArgsKwargs(25, ratio=(0.5, 1.5)),
             ArgsKwargs((31, 28), interpolation=v2_transforms.InterpolationMode.NEAREST),
-            ArgsKwargs((33, 26), interpolation=v2_transforms.InterpolationMode.BICUBIC),
             ArgsKwargs((31, 28), interpolation=PIL.Image.NEAREST),
-            ArgsKwargs((33, 26), interpolation=PIL.Image.BICUBIC),
             ArgsKwargs((29, 32), antialias=False),
             ArgsKwargs((28, 31), antialias=True),
         ],
         # atol=1 due to Resize v2 is using native uint8 interpolate path for bilinear and nearest modes
         closeness_kwargs=dict(rtol=0, atol=1),
     ),
+    ConsistencyConfig(
+        v2_transforms.RandomResizedCrop,
+        legacy_transforms.RandomResizedCrop,
+        [
+            ArgsKwargs((33, 26), interpolation=v2_transforms.InterpolationMode.BICUBIC, antialias=True),
+            ArgsKwargs((33, 26), interpolation=PIL.Image.BICUBIC, antialias=True),
+        ],
+        closeness_kwargs=dict(rtol=0, atol=21),
+    ),
     ConsistencyConfig(
         v2_transforms.RandomErasing,
         legacy_transforms.RandomErasing,

test/transforms_v2_kernel_infos.py

@@ -257,17 +257,20 @@ def sample_inputs_resize_image_tensor():
 
     for image_loader, interpolation in itertools.product(
         make_image_loaders(sizes=["random"], color_spaces=["RGB"]),
-        [
-            F.InterpolationMode.NEAREST,
-            F.InterpolationMode.BILINEAR,
-            F.InterpolationMode.BICUBIC,
-        ],
+        [F.InterpolationMode.NEAREST, F.InterpolationMode.BILINEAR],
     ):
         yield ArgsKwargs(image_loader, size=[min(image_loader.spatial_size) + 1], interpolation=interpolation)
 
     yield ArgsKwargs(make_image_loader(size=(11, 17)), size=20, max_size=25)
 
 
+def sample_inputs_resize_image_tensor_bicubic():
+    for image_loader, interpolation in itertools.product(
+        make_image_loaders(sizes=["random"], color_spaces=["RGB"]), [F.InterpolationMode.BICUBIC]
+    ):
+        yield ArgsKwargs(image_loader, size=[min(image_loader.spatial_size) + 1], interpolation=interpolation)
+
+
 @pil_reference_wrapper
 def reference_resize_image_tensor(*args, **kwargs):
     if not kwargs.pop("antialias", False) and kwargs.get("interpolation", F.InterpolationMode.BILINEAR) in {
@@ -364,6 +367,21 @@ KERNEL_INFOS.extend(
                 xfail_jit_python_scalar_arg("size"),
             ],
         ),
+        KernelInfo(
+            F.resize_image_tensor,
+            sample_inputs_fn=sample_inputs_resize_image_tensor_bicubic,
+            reference_fn=reference_resize_image_tensor,
+            reference_inputs_fn=reference_inputs_resize_image_tensor,
+            float32_vs_uint8=True,
+            closeness_kwargs={
+                **pil_reference_pixel_difference(10, mae=True),
+                **cuda_vs_cpu_pixel_difference(atol=30),
+                **float32_vs_uint8_pixel_difference(1, mae=True),
+            },
+            test_marks=[
+                xfail_jit_python_scalar_arg("size"),
+            ],
+        ),
         KernelInfo(
             F.resize_bounding_box,
             sample_inputs_fn=sample_inputs_resize_bounding_box,

torchvision/transforms/v2/functional/_geometry.py

@@ -190,14 +190,13 @@ def resize_image_tensor(
         if interpolation == InterpolationMode.NEAREST or interpolation == InterpolationMode.NEAREST_EXACT:
             # uint8 dtype can be included for cpu and cuda input if nearest mode
             acceptable_dtypes.append(torch.uint8)
-        elif (
-            interpolation == InterpolationMode.BILINEAR
-            and image.device.type == "cpu"
-            and "AVX2" in torch.backends.cpu.get_cpu_capability()
-        ):
-            # uint8 dtype support for bilinear mode is limited to cpu and
-            # according to our benchmarks non-AVX CPUs should prefer u8->f32->interpolate->u8 path
-            acceptable_dtypes.append(torch.uint8)
+        elif image.device.type == "cpu":
+            # uint8 dtype support for bilinear and bicubic is limited to cpu and
+            # according to our benchmarks, non-AVX CPUs should still prefer u8->f32->interpolate->u8 path for bilinear
+            if (interpolation == InterpolationMode.BILINEAR and "AVX2" in torch.backends.cpu.get_cpu_capability()) or (
+                interpolation == InterpolationMode.BICUBIC
+            ):
+                acceptable_dtypes.append(torch.uint8)
 
         strides = image.stride()
         if image.is_contiguous(memory_format=torch.channels_last) and image.shape[0] == 1 and numel != strides[0]:
@@ -227,6 +226,7 @@ def resize_image_tensor(
 
         if need_cast:
             if interpolation == InterpolationMode.BICUBIC and dtype == torch.uint8:
+                # This path is hit on non-AVX archs, or on GPU.
                 image = image.clamp_(min=0, max=255)
             if dtype in (torch.uint8, torch.int8, torch.int16, torch.int32, torch.int64):
                 image = image.round_()