Perspective CV-CUDA Backend

test/test_transforms_v2.py

@@ -21,6 +21,7 @@
 import torchvision.transforms.v2 as transforms
 
 from common_utils import (
+    assert_close,
     assert_equal,
     cache,
     cpu_and_cuda,
@@ -41,7 +42,6 @@
 )
 
 from torch import nn
-from torch.testing import assert_close
 from torch.utils._pytree import tree_flatten, tree_map
 from torch.utils.data import DataLoader, default_collate
 from torchvision import tv_tensors
@@ -5180,6 +5180,9 @@ def test_kernel_video(self):
             make_segmentation_mask,
             make_video,
             make_keypoints,
+            pytest.param(
+                make_image_cvcuda, marks=pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="CVCUDA not available")
+            ),
         ],
     )
     def test_functional(self, make_input):
@@ -5195,9 +5198,16 @@ def test_functional(self, make_input):
             (F.perspective_mask, tv_tensors.Mask),
             (F.perspective_video, tv_tensors.Video),
             (F.perspective_keypoints, tv_tensors.KeyPoints),
+            pytest.param(
+                F._geometry._perspective_image_cvcuda,
+                None,
+                marks=pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="CVCUDA not available"),
+            ),
         ],
     )
     def test_functional_signature(self, kernel, input_type):
+        if kernel is F._geometry._perspective_image_cvcuda:
+            input_type = _import_cvcuda().Tensor
         check_functional_kernel_signature_match(F.perspective, kernel=kernel, input_type=input_type)
 
     @pytest.mark.parametrize("distortion_scale", [0.5, 0.0, 1.0])
@@ -5211,6 +5221,9 @@ def test_functional_signature(self, kernel, input_type):
             make_segmentation_mask,
             make_video,
             make_keypoints,
+            pytest.param(
+                make_image_cvcuda, marks=pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="CVCUDA not available")
+            ),
         ],
     )
     def test_transform(self, distortion_scale, make_input):
@@ -5226,12 +5239,25 @@ def test_transform_error(self, distortion_scale):
         "interpolation", [transforms.InterpolationMode.NEAREST, transforms.InterpolationMode.BILINEAR]
     )
     @pytest.mark.parametrize("fill", CORRECTNESS_FILLS)
-    def test_image_functional_correctness(self, coefficients, interpolation, fill):
-        image = make_image(dtype=torch.uint8, device="cpu")
+    @pytest.mark.parametrize(
+        "make_input",
+        [
+            make_image,
+            pytest.param(
+                make_image_cvcuda, marks=pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="CVCUDA not available")
+            ),
+        ],
+    )
+    def test_image_functional_correctness(self, coefficients, interpolation, fill, make_input):
+        image = make_input(dtype=torch.uint8, device="cpu")
 
         actual = F.perspective(
             image, startpoints=None, endpoints=None, coefficients=coefficients, interpolation=interpolation, fill=fill
         )
+        if make_input is make_image_cvcuda:
+            actual = F.cvcuda_to_tensor(actual)[0].cpu()
+            image = F.cvcuda_to_tensor(image)[0].cpu()
+
         expected = F.to_image(
             F.perspective(
                 F.to_pil_image(image),
@@ -5243,13 +5269,20 @@ def test_image_functional_correctness(self, coefficients, interpolation, fill):
             )
         )
 
-        if interpolation is transforms.InterpolationMode.BILINEAR:
-            abs_diff = (actual.float() - expected.float()).abs()
-            assert (abs_diff > 1).float().mean() < 7e-2
-            mae = abs_diff.mean()
-            assert mae < 3
-        else:
-            assert_equal(actual, expected)
+        if make_input is make_image:
+            if interpolation is transforms.InterpolationMode.BILINEAR:
+                abs_diff = (actual.float() - expected.float()).abs()
+                assert (abs_diff > 1).float().mean() < 7e-2
+                mae = abs_diff.mean()
+                assert mae < 3
+            else:
+                assert_equal(actual, expected)
+        else:  # CV-CUDA
+            # just check that the shapes/dtypes are the same, cvcuda warp_perspective uses different algorithm
+            # visually the results are the same on real images,
+            # realistically, the diff is not visible to the human eye
+            tolerance = 255 if interpolation is transforms.InterpolationMode.NEAREST else 125
+            assert_close(actual, expected, rtol=0, atol=tolerance)
 
     def _reference_perspective_bounding_boxes(self, bounding_boxes, *, startpoints, endpoints):
         format = bounding_boxes.format

torchvision/transforms/v2/_geometry.py

@@ -944,6 +944,9 @@ class RandomPerspective(_RandomApplyTransform):
 
     _v1_transform_cls = _transforms.RandomPerspective
 
+    if CVCUDA_AVAILABLE:
+        _transformed_types = _RandomApplyTransform._transformed_types + (_is_cvcuda_tensor,)
+
     def __init__(
         self,
         distortion_scale: float = 0.5,

torchvision/transforms/v2/_utils.py

@@ -16,7 +16,7 @@
 
 from torchvision.transforms.transforms import _check_sequence_input, _setup_angle, _setup_size  # noqa: F401
 from torchvision.transforms.v2.functional import get_dimensions, get_size, is_pure_tensor
-from torchvision.transforms.v2.functional._utils import _FillType, _FillTypeJIT
+from torchvision.transforms.v2.functional._utils import _FillType, _FillTypeJIT, _is_cvcuda_tensor
 
 
 def _setup_number_or_seq(arg: int | float | Sequence[int | float], name: str) -> Sequence[float]:
@@ -182,7 +182,7 @@ def query_chw(flat_inputs: list[Any]) -> tuple[int, int, int]:
     chws = {
         tuple(get_dimensions(inpt))
         for inpt in flat_inputs
-        if check_type(inpt, (is_pure_tensor, tv_tensors.Image, PIL.Image.Image, tv_tensors.Video))
+        if check_type(inpt, (is_pure_tensor, tv_tensors.Image, PIL.Image.Image, tv_tensors.Video, _is_cvcuda_tensor))
     }
     if not chws:
         raise TypeError("No image or video was found in the sample")
@@ -207,6 +207,7 @@ def query_size(flat_inputs: list[Any]) -> tuple[int, int]:
                 tv_tensors.Mask,
                 tv_tensors.BoundingBoxes,
                 tv_tensors.KeyPoints,
+                _is_cvcuda_tensor,
             ),
         )
     }

torchvision/transforms/v2/functional/_geometry.py

@@ -4,6 +4,8 @@
 from collections.abc import Sequence
 from typing import Any, Optional, TYPE_CHECKING, Union
 
+import numpy as np
+
 import PIL.Image
 import torch
 from torch.nn.functional import grid_sample, interpolate, pad as torch_pad
@@ -28,6 +30,7 @@
 
 from ._utils import (
     _FillTypeJIT,
+    _get_cvcuda_interp,
     _get_kernel,
     _import_cvcuda,
     _is_cvcuda_available,
@@ -2285,6 +2288,44 @@ def perspective_video(
     )
 
 
+def _perspective_image_cvcuda(
+    image: "cvcuda.Tensor",
+    startpoints: Optional[list[list[int]]],
+    endpoints: Optional[list[list[int]]],
+    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
+    fill: _FillTypeJIT = None,
+    coefficients: Optional[list[float]] = None,
+) -> "cvcuda.Tensor":
+    cvcuda = _import_cvcuda()
+
+    c = _perspective_coefficients(startpoints, endpoints, coefficients)
+    interpolation = _check_interpolation(interpolation)
+
+    interp = _get_cvcuda_interp(interpolation)
+
+    xform = np.array([[c[0], c[1], c[2]], [c[3], c[4], c[5]], [c[6], c[7], 1.0]], dtype=np.float32)
+
+    num_channels = image.shape[-1]
+    if fill is None:
+        border_value = np.zeros(num_channels, dtype=np.float32)
+    elif isinstance(fill, (int, float)):
+        border_value = np.full(num_channels, fill, dtype=np.float32)
+    else:
+        border_value = np.array(fill, dtype=np.float32)[:num_channels]
+
+    return cvcuda.warp_perspective(
+        image,
+        xform,
+        flags=interp | cvcuda.Interp.WARP_INVERSE_MAP,
+        border_mode=cvcuda.Border.CONSTANT,
+        border_value=border_value,
+    )
+
+
+if CVCUDA_AVAILABLE:
+    _register_kernel_internal(perspective, _import_cvcuda().Tensor)(_perspective_image_cvcuda)
+
+
 def elastic(
     inpt: torch.Tensor,
     displacement: torch.Tensor,

torchvision/transforms/v2/functional/_meta.py

@@ -51,6 +51,16 @@ def get_dimensions_video(video: torch.Tensor) -> list[int]:
     return get_dimensions_image(video)
 
 
+def get_dimensions_image_cvcuda(image: "cvcuda.Tensor") -> list[int]:
+    # CV-CUDA tensor is always in NHWC layout
+    # get_dimensions is CHW
+    return [image.shape[3], image.shape[1], image.shape[2]]
+
+
+if CVCUDA_AVAILABLE:
+    _register_kernel_internal(get_dimensions, cvcuda.Tensor)(get_dimensions_image_cvcuda)
+
+
 def get_num_channels(inpt: torch.Tensor) -> int:
     if torch.jit.is_scripting():
         return get_num_channels_image(inpt)
@@ -87,6 +97,16 @@ def get_num_channels_video(video: torch.Tensor) -> int:
 get_image_num_channels = get_num_channels
 
 
+def get_num_channels_image_cvcuda(image: "cvcuda.Tensor") -> int:
+    # CV-CUDA tensor is always in NHWC layout
+    # get_num_channels is C
+    return image.shape[3]
+
+
+if CVCUDA_AVAILABLE:
+    _register_kernel_internal(get_num_channels, cvcuda.Tensor)(get_num_channels_image_cvcuda)
+
+
 def get_size(inpt: torch.Tensor) -> list[int]:
     if torch.jit.is_scripting():
         return get_size_image(inpt)
@@ -125,7 +145,7 @@ def get_size_image_cvcuda(image: "cvcuda.Tensor") -> list[int]:
 
 
 if CVCUDA_AVAILABLE:
-    _get_size_image_cvcuda = _register_kernel_internal(get_size, cvcuda.Tensor)(get_size_image_cvcuda)
+    _register_kernel_internal(get_size, _import_cvcuda().Tensor)(get_size_image_cvcuda)
 
 
 @_register_kernel_internal(get_size, tv_tensors.Video, tv_tensor_wrapper=False)

torchvision/transforms/v2/functional/_utils.py

@@ -1,9 +1,13 @@
 import functools
 from collections.abc import Sequence
-from typing import Any, Callable, Optional, Union
+from typing import Any, Callable, Optional, TYPE_CHECKING, Union
 
 import torch
 from torchvision import tv_tensors
+from torchvision.transforms.functional import InterpolationMode
+
+if TYPE_CHECKING:
+    import cvcuda  # type: ignore[import-not-found]
 
 _FillType = Union[int, float, Sequence[int], Sequence[float], None]
 _FillTypeJIT = Optional[list[float]]
@@ -177,3 +181,37 @@ def _is_cvcuda_tensor(inpt: Any) -> bool:
         return isinstance(inpt, cvcuda.Tensor)
     except ImportError:
         return False
+
+
+_interpolation_mode_to_cvcuda_interp: dict[InterpolationMode | str | int, "cvcuda.Interp"] = {}
+
+
+def _get_cvcuda_interp(interpolation: InterpolationMode | str | int) -> "cvcuda.Interp":
+    if len(_interpolation_mode_to_cvcuda_interp) == 0:
+        cvcuda = _import_cvcuda()
+        _interpolation_mode_to_cvcuda_interp[InterpolationMode.NEAREST] = cvcuda.Interp.NEAREST
+        _interpolation_mode_to_cvcuda_interp[InterpolationMode.NEAREST_EXACT] = cvcuda.Interp.NEAREST
+        _interpolation_mode_to_cvcuda_interp[InterpolationMode.BILINEAR] = cvcuda.Interp.LINEAR
+        _interpolation_mode_to_cvcuda_interp[InterpolationMode.BICUBIC] = cvcuda.Interp.CUBIC
+        _interpolation_mode_to_cvcuda_interp[InterpolationMode.BOX] = cvcuda.Interp.BOX
+        _interpolation_mode_to_cvcuda_interp[InterpolationMode.HAMMING] = cvcuda.Interp.HAMMING
+        _interpolation_mode_to_cvcuda_interp[InterpolationMode.LANCZOS] = cvcuda.Interp.LANCZOS
+        _interpolation_mode_to_cvcuda_interp["nearest"] = cvcuda.Interp.NEAREST
+        _interpolation_mode_to_cvcuda_interp["nearest-exact"] = cvcuda.Interp.NEAREST
+        _interpolation_mode_to_cvcuda_interp["bilinear"] = cvcuda.Interp.LINEAR
+        _interpolation_mode_to_cvcuda_interp["bicubic"] = cvcuda.Interp.CUBIC
+        _interpolation_mode_to_cvcuda_interp["box"] = cvcuda.Interp.BOX
+        _interpolation_mode_to_cvcuda_interp["hamming"] = cvcuda.Interp.HAMMING
+        _interpolation_mode_to_cvcuda_interp["lanczos"] = cvcuda.Interp.LANCZOS
+        _interpolation_mode_to_cvcuda_interp[0] = cvcuda.Interp.NEAREST
+        _interpolation_mode_to_cvcuda_interp[2] = cvcuda.Interp.LINEAR
+        _interpolation_mode_to_cvcuda_interp[3] = cvcuda.Interp.CUBIC
+        _interpolation_mode_to_cvcuda_interp[4] = cvcuda.Interp.BOX
+        _interpolation_mode_to_cvcuda_interp[5] = cvcuda.Interp.HAMMING
+        _interpolation_mode_to_cvcuda_interp[1] = cvcuda.Interp.LANCZOS
+
+    interp = _interpolation_mode_to_cvcuda_interp.get(interpolation)
+    if interp is None:
+        raise ValueError(f"Interpolation mode {interpolation} is not supported with CV-CUDA")
+
+    return interp