From 44db71c0772e5ef5758c38d0e4e8ad9995946c80 Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Tue, 25 Nov 2025 09:14:49 -0800 Subject: [PATCH 01/15] implement additional cvcuda infra for all branches to avoid duplicate setup --- torchvision/transforms/v2/_transform.py | 4 ++-- torchvision/transforms/v2/_utils.py | 3 ++- .../transforms/v2/functional/__init__.py | 2 +- .../transforms/v2/functional/_augment.py | 11 ++++++++++- .../transforms/v2/functional/_color.py | 12 +++++++++++- .../transforms/v2/functional/_geometry.py | 19 +++++++++++++++++-- torchvision/transforms/v2/functional/_misc.py | 11 +++++++++-- .../transforms/v2/functional/_utils.py | 16 ++++++++++++++++ 8 files changed, 68 insertions(+), 10 deletions(-) diff --git a/torchvision/transforms/v2/_transform.py b/torchvision/transforms/v2/_transform.py index ac84fcb6c82..bec9ffcf714 100644 --- a/torchvision/transforms/v2/_transform.py +++ b/torchvision/transforms/v2/_transform.py @@ -11,7 +11,7 @@ from torchvision.transforms.v2._utils import check_type, has_any, is_pure_tensor from torchvision.utils import _log_api_usage_once -from .functional._utils import _get_kernel +from .functional._utils import _get_kernel, is_cvcuda_tensor class Transform(nn.Module): @@ -23,7 +23,7 @@ class Transform(nn.Module): # Class attribute defining transformed types. Other types are passed-through without any transformation # We support both Types and callables that are able to do further checks on the type of the input. - _transformed_types: tuple[type | Callable[[Any], bool], ...] = (torch.Tensor, PIL.Image.Image) + _transformed_types: tuple[type | Callable[[Any], bool], ...] = (torch.Tensor, PIL.Image.Image, is_cvcuda_tensor) def __init__(self) -> None: super().__init__() diff --git a/torchvision/transforms/v2/_utils.py b/torchvision/transforms/v2/_utils.py index bb6051b4e61..765a772fe41 100644 --- a/torchvision/transforms/v2/_utils.py +++ b/torchvision/transforms/v2/_utils.py @@ -15,7 +15,7 @@ from torchvision._utils import sequence_to_str 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 import get_dimensions, get_size, is_cvcuda_tensor, is_pure_tensor from torchvision.transforms.v2.functional._utils import _FillType, _FillTypeJIT @@ -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, ), ) } diff --git a/torchvision/transforms/v2/functional/__init__.py b/torchvision/transforms/v2/functional/__init__.py index 032a993b1f0..52181e4624b 100644 --- a/torchvision/transforms/v2/functional/__init__.py +++ b/torchvision/transforms/v2/functional/__init__.py @@ -1,6 +1,6 @@ from torchvision.transforms import InterpolationMode # usort: skip -from ._utils import is_pure_tensor, register_kernel # usort: skip +from ._utils import is_pure_tensor, register_kernel, is_cvcuda_tensor # usort: skip from ._meta import ( clamp_bounding_boxes, diff --git a/torchvision/transforms/v2/functional/_augment.py b/torchvision/transforms/v2/functional/_augment.py index a904d8d7cbd..7ce5bdc7b7e 100644 --- a/torchvision/transforms/v2/functional/_augment.py +++ b/torchvision/transforms/v2/functional/_augment.py @@ -1,4 +1,5 @@ import io +from typing import TYPE_CHECKING import PIL.Image @@ -8,7 +9,15 @@ from torchvision.transforms.functional import pil_to_tensor, to_pil_image from torchvision.utils import _log_api_usage_once -from ._utils import _get_kernel, _register_kernel_internal +from ._utils import _get_kernel, _import_cvcuda, _is_cvcuda_available, _register_kernel_internal + + +CVCUDA_AVAILABLE = _is_cvcuda_available() + +if TYPE_CHECKING: + import cvcuda # type: ignore[import-not-found] +if CVCUDA_AVAILABLE: + cvcuda = _import_cvcuda() # noqa: F811 def erase( diff --git a/torchvision/transforms/v2/functional/_color.py b/torchvision/transforms/v2/functional/_color.py index be254c0d63a..5be9c62902a 100644 --- a/torchvision/transforms/v2/functional/_color.py +++ b/torchvision/transforms/v2/functional/_color.py @@ -1,3 +1,5 @@ +from typing import TYPE_CHECKING + import PIL.Image import torch from torch.nn.functional import conv2d @@ -9,7 +11,15 @@ from ._misc import _num_value_bits, to_dtype_image from ._type_conversion import pil_to_tensor, to_pil_image -from ._utils import _get_kernel, _register_kernel_internal +from ._utils import _get_kernel, _import_cvcuda, _is_cvcuda_available, _register_kernel_internal + + +CVCUDA_AVAILABLE = _is_cvcuda_available() + +if TYPE_CHECKING: + import cvcuda # type: ignore[import-not-found] +if CVCUDA_AVAILABLE: + cvcuda = _import_cvcuda() # noqa: F811 def rgb_to_grayscale(inpt: torch.Tensor, num_output_channels: int = 1) -> torch.Tensor: diff --git a/torchvision/transforms/v2/functional/_geometry.py b/torchvision/transforms/v2/functional/_geometry.py index 4fcb7fabe0d..c029488001c 100644 --- a/torchvision/transforms/v2/functional/_geometry.py +++ b/torchvision/transforms/v2/functional/_geometry.py @@ -2,7 +2,7 @@ import numbers import warnings from collections.abc import Sequence -from typing import Any, Optional, Union +from typing import Any, Optional, TYPE_CHECKING, Union import PIL.Image import torch @@ -26,7 +26,22 @@ from ._meta import _get_size_image_pil, clamp_bounding_boxes, convert_bounding_box_format -from ._utils import _FillTypeJIT, _get_kernel, _register_five_ten_crop_kernel_internal, _register_kernel_internal +from ._utils import ( + _FillTypeJIT, + _get_kernel, + _import_cvcuda, + _is_cvcuda_available, + _register_five_ten_crop_kernel_internal, + _register_kernel_internal, +) + + +CVCUDA_AVAILABLE = _is_cvcuda_available() + +if TYPE_CHECKING: + import cvcuda # type: ignore[import-not-found] +if CVCUDA_AVAILABLE: + cvcuda = _import_cvcuda() # noqa: F811 def _check_interpolation(interpolation: Union[InterpolationMode, int]) -> InterpolationMode: diff --git a/torchvision/transforms/v2/functional/_misc.py b/torchvision/transforms/v2/functional/_misc.py index daf263df046..0fa05a2113c 100644 --- a/torchvision/transforms/v2/functional/_misc.py +++ b/torchvision/transforms/v2/functional/_misc.py @@ -1,5 +1,5 @@ import math -from typing import Optional +from typing import Optional, TYPE_CHECKING import PIL.Image import torch @@ -13,7 +13,14 @@ from ._meta import _convert_bounding_box_format -from ._utils import _get_kernel, _register_kernel_internal, is_pure_tensor +from ._utils import _get_kernel, _import_cvcuda, _is_cvcuda_available, _register_kernel_internal, is_pure_tensor + +CVCUDA_AVAILABLE = _is_cvcuda_available() + +if TYPE_CHECKING: + import cvcuda # type: ignore[import-not-found] +if CVCUDA_AVAILABLE: + cvcuda = _import_cvcuda() # noqa: F811 def normalize( diff --git a/torchvision/transforms/v2/functional/_utils.py b/torchvision/transforms/v2/functional/_utils.py index ad1eddd258b..73fafaf7425 100644 --- a/torchvision/transforms/v2/functional/_utils.py +++ b/torchvision/transforms/v2/functional/_utils.py @@ -169,3 +169,19 @@ def _is_cvcuda_available(): return True except ImportError: return False + + +def is_cvcuda_tensor(inpt: Any) -> bool: + """ + Check if the input is a CVCUDA tensor. + + Args: + inpt: The input to check. + + Returns: + True if the input is a CV-CUDA tensor, False otherwise. + """ + if _is_cvcuda_available(): + cvcuda = _import_cvcuda() + return isinstance(inpt, cvcuda.Tensor) + return False From e3dd70022fa1c87aca7a9a98068b6e13e802a375 Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Tue, 25 Nov 2025 09:26:19 -0800 Subject: [PATCH 02/15] update make_image_cvcuda to have default batch dim --- test/common_utils.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/test/common_utils.py b/test/common_utils.py index 8c3c9dd58a8..e7bae60c41b 100644 --- a/test/common_utils.py +++ b/test/common_utils.py @@ -400,8 +400,9 @@ def make_image_pil(*args, **kwargs): return to_pil_image(make_image(*args, **kwargs)) -def make_image_cvcuda(*args, **kwargs): - return to_cvcuda_tensor(make_image(*args, **kwargs)) +def make_image_cvcuda(*args, batch_dims=(1,), **kwargs): + # explicitly default batch_dims to (1,) since to_cvcuda_tensor requires a batch dimension (ndims == 4) + return to_cvcuda_tensor(make_image(*args, batch_dims=batch_dims, **kwargs)) def make_keypoints(canvas_size=DEFAULT_SIZE, *, num_points=4, dtype=None, device="cpu"): From c035df1c6eaebcad25604f8c298a7d9eaf86864b Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Mon, 1 Dec 2025 18:16:27 -0800 Subject: [PATCH 03/15] add stanardized setup to main for easier updating of PRs and branches --- test/common_utils.py | 21 ++++++++++++++-- test/test_transforms_v2.py | 2 +- torchvision/transforms/v2/_utils.py | 2 +- torchvision/transforms/v2/functional/_meta.py | 24 +++++++++++++++++-- 4 files changed, 43 insertions(+), 6 deletions(-) diff --git a/test/common_utils.py b/test/common_utils.py index e7bae60c41b..3b889e93d2e 100644 --- a/test/common_utils.py +++ b/test/common_utils.py @@ -20,13 +20,15 @@ from torch.testing._comparison import BooleanPair, NonePair, not_close_error_metas, NumberPair, TensorLikePair from torchvision import io, tv_tensors from torchvision.transforms._functional_tensor import _max_value as get_max_value -from torchvision.transforms.v2.functional import to_cvcuda_tensor, to_image, to_pil_image +from torchvision.transforms.v2.functional import cvcuda_to_tensor, to_cvcuda_tensor, to_image, to_pil_image +from torchvision.transforms.v2.functional._utils import _import_cvcuda, _is_cvcuda_available from torchvision.utils import _Image_fromarray IN_OSS_CI = any(os.getenv(var) == "true" for var in ["CIRCLECI", "GITHUB_ACTIONS"]) IN_RE_WORKER = os.environ.get("INSIDE_RE_WORKER") is not None IN_FBCODE = os.environ.get("IN_FBCODE_TORCHVISION") == "1" +CVCUDA_AVAILABLE = _is_cvcuda_available() CUDA_NOT_AVAILABLE_MSG = "CUDA device not available" MPS_NOT_AVAILABLE_MSG = "MPS device not available" OSS_CI_GPU_NO_CUDA_MSG = "We're in an OSS GPU machine, and this test doesn't need cuda." @@ -275,6 +277,17 @@ def combinations_grid(**kwargs): return [dict(zip(kwargs.keys(), values)) for values in itertools.product(*kwargs.values())] +def cvcuda_to_pil_compatible_tensor(tensor: "cvcuda.Tensor") -> torch.Tensor: + tensor = cvcuda_to_tensor(tensor) + if tensor.ndim != 4: + raise ValueError(f"CV-CUDA Tensor should be 4 dimensional. Got {tensor.ndim} dimensions.") + if tensor.shape[0] != 1: + raise ValueError( + f"CV-CUDA Tensor should have batch dimension 1 for comparison with PIL.Image.Image. Got {tensor.shape[0]}." + ) + return tensor.squeeze(0).cpu() + + class ImagePair(TensorLikePair): def __init__( self, @@ -287,6 +300,11 @@ def __init__( if all(isinstance(input, PIL.Image.Image) for input in [actual, expected]): actual, expected = (to_image(input) for input in [actual, expected]) + # handle check for CV-CUDA Tensors + if CVCUDA_AVAILABLE and isinstance(actual, _import_cvcuda().Tensor): + # Use the PIL compatible tensor, so we can always compare with PIL.Image.Image + actual = cvcuda_to_pil_compatible_tensor(actual) + super().__init__(actual, expected, **other_parameters) self.mae = mae @@ -401,7 +419,6 @@ def make_image_pil(*args, **kwargs): def make_image_cvcuda(*args, batch_dims=(1,), **kwargs): - # explicitly default batch_dims to (1,) since to_cvcuda_tensor requires a batch dimension (ndims == 4) return to_cvcuda_tensor(make_image(*args, batch_dims=batch_dims, **kwargs)) diff --git a/test/test_transforms_v2.py b/test/test_transforms_v2.py index 670a9d00ffb..7eba65550da 100644 --- a/test/test_transforms_v2.py +++ b/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 diff --git a/torchvision/transforms/v2/_utils.py b/torchvision/transforms/v2/_utils.py index 765a772fe41..3fc33ce5964 100644 --- a/torchvision/transforms/v2/_utils.py +++ b/torchvision/transforms/v2/_utils.py @@ -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") diff --git a/torchvision/transforms/v2/functional/_meta.py b/torchvision/transforms/v2/functional/_meta.py index 6b8f19f12f4..ee562cb2aee 100644 --- a/torchvision/transforms/v2/functional/_meta.py +++ b/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_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_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_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_cvcuda) + + def get_size(inpt: torch.Tensor) -> list[int]: if torch.jit.is_scripting(): return get_size_image(inpt) @@ -114,7 +134,7 @@ def _get_size_image_pil(image: PIL.Image.Image) -> list[int]: return [height, width] -def get_size_image_cvcuda(image: "cvcuda.Tensor") -> list[int]: +def _get_size_cvcuda(image: "cvcuda.Tensor") -> list[int]: """Get size of `cvcuda.Tensor` with NHWC layout.""" hw = list(image.shape[-3:-1]) ndims = len(hw) @@ -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, cvcuda.Tensor)(_get_size_cvcuda) @_register_kernel_internal(get_size, tv_tensors.Video, tv_tensor_wrapper=False) From 98d7dfb2059eaf2c10c3f549ea45f1d27875134c Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Mon, 1 Dec 2025 18:25:09 -0800 Subject: [PATCH 04/15] update is_cvcuda_tensor --- torchvision/transforms/v2/functional/_utils.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/torchvision/transforms/v2/functional/_utils.py b/torchvision/transforms/v2/functional/_utils.py index 73fafaf7425..44b2edeaf2d 100644 --- a/torchvision/transforms/v2/functional/_utils.py +++ b/torchvision/transforms/v2/functional/_utils.py @@ -181,7 +181,8 @@ def is_cvcuda_tensor(inpt: Any) -> bool: Returns: True if the input is a CV-CUDA tensor, False otherwise. """ - if _is_cvcuda_available(): + try: cvcuda = _import_cvcuda() return isinstance(inpt, cvcuda.Tensor) - return False + except ImportError: + return False From ddc116d13febdae1d53507bcde9f103a4c14eba7 Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Tue, 2 Dec 2025 12:37:03 -0800 Subject: [PATCH 05/15] add cvcuda to pil compatible to transforms by default --- test/test_transforms_v2.py | 1 + 1 file changed, 1 insertion(+) diff --git a/test/test_transforms_v2.py b/test/test_transforms_v2.py index 7eba65550da..87166477669 100644 --- a/test/test_transforms_v2.py +++ b/test/test_transforms_v2.py @@ -25,6 +25,7 @@ assert_equal, cache, cpu_and_cuda, + cvcuda_to_pil_compatible_tensor, freeze_rng_state, ignore_jit_no_profile_information_warning, make_bounding_boxes, From e51dc7eabd254261347245f4492892fd0944aae5 Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Tue, 2 Dec 2025 12:46:23 -0800 Subject: [PATCH 06/15] remove cvcuda from transform class --- torchvision/transforms/v2/_transform.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/torchvision/transforms/v2/_transform.py b/torchvision/transforms/v2/_transform.py index bec9ffcf714..ac84fcb6c82 100644 --- a/torchvision/transforms/v2/_transform.py +++ b/torchvision/transforms/v2/_transform.py @@ -11,7 +11,7 @@ from torchvision.transforms.v2._utils import check_type, has_any, is_pure_tensor from torchvision.utils import _log_api_usage_once -from .functional._utils import _get_kernel, is_cvcuda_tensor +from .functional._utils import _get_kernel class Transform(nn.Module): @@ -23,7 +23,7 @@ class Transform(nn.Module): # Class attribute defining transformed types. Other types are passed-through without any transformation # We support both Types and callables that are able to do further checks on the type of the input. - _transformed_types: tuple[type | Callable[[Any], bool], ...] = (torch.Tensor, PIL.Image.Image, is_cvcuda_tensor) + _transformed_types: tuple[type | Callable[[Any], bool], ...] = (torch.Tensor, PIL.Image.Image) def __init__(self) -> None: super().__init__() From 4939355a2c7421eeba95d7f155fe7953066aec6d Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Thu, 4 Dec 2025 11:07:08 -0800 Subject: [PATCH 07/15] resolve more formatting naming --- torchvision/transforms/v2/functional/__init__.py | 2 +- torchvision/transforms/v2/functional/_meta.py | 8 ++++---- 2 files changed, 5 insertions(+), 5 deletions(-) diff --git a/torchvision/transforms/v2/functional/__init__.py b/torchvision/transforms/v2/functional/__init__.py index 52181e4624b..032a993b1f0 100644 --- a/torchvision/transforms/v2/functional/__init__.py +++ b/torchvision/transforms/v2/functional/__init__.py @@ -1,6 +1,6 @@ from torchvision.transforms import InterpolationMode # usort: skip -from ._utils import is_pure_tensor, register_kernel, is_cvcuda_tensor # usort: skip +from ._utils import is_pure_tensor, register_kernel # usort: skip from ._meta import ( clamp_bounding_boxes, diff --git a/torchvision/transforms/v2/functional/_meta.py b/torchvision/transforms/v2/functional/_meta.py index e8630f788ca..af03ad018d4 100644 --- a/torchvision/transforms/v2/functional/_meta.py +++ b/torchvision/transforms/v2/functional/_meta.py @@ -51,14 +51,14 @@ def get_dimensions_video(video: torch.Tensor) -> list[int]: return get_dimensions_image(video) -def _get_dimensions_cvcuda(image: "cvcuda.Tensor") -> list[int]: +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_cvcuda) + _register_kernel_internal(get_dimensions, cvcuda.Tensor)(get_dimensions_image_cvcuda) def get_num_channels(inpt: torch.Tensor) -> int: @@ -97,14 +97,14 @@ def get_num_channels_video(video: torch.Tensor) -> int: get_image_num_channels = get_num_channels -def _get_num_channels_cvcuda(image: "cvcuda.Tensor") -> int: +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_cvcuda) + _register_kernel_internal(get_num_channels, cvcuda.Tensor)(get_num_channels_image_cvcuda) def get_size(inpt: torch.Tensor) -> list[int]: From fbea584365311ae6b56be7e4f6bbff1f834dd31a Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Thu, 4 Dec 2025 11:15:49 -0800 Subject: [PATCH 08/15] update is cvcuda tensor impl --- torchvision/transforms/v2/_utils.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/torchvision/transforms/v2/_utils.py b/torchvision/transforms/v2/_utils.py index 3fc33ce5964..e803aa49c60 100644 --- a/torchvision/transforms/v2/_utils.py +++ b/torchvision/transforms/v2/_utils.py @@ -15,8 +15,8 @@ from torchvision._utils import sequence_to_str 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_cvcuda_tensor, is_pure_tensor -from torchvision.transforms.v2.functional._utils import _FillType, _FillTypeJIT +from torchvision.transforms.v2.functional import get_dimensions, get_size, is_pure_tensor +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, is_cvcuda_tensor)) + 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") @@ -207,7 +207,7 @@ def query_size(flat_inputs: list[Any]) -> tuple[int, int]: tv_tensors.Mask, tv_tensors.BoundingBoxes, tv_tensors.KeyPoints, - is_cvcuda_tensor, + _is_cvcuda_tensor, ), ) } From f8aab311dfebcf40bc5196444e68996e24f7d371 Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Wed, 26 Nov 2025 14:09:35 -0800 Subject: [PATCH 09/15] wip elastic --- test/test_transforms_v2.py | 16 +++ .../transforms/v2/functional/_geometry.py | 106 ++++++++++++++++++ 2 files changed, 122 insertions(+) diff --git a/test/test_transforms_v2.py b/test/test_transforms_v2.py index f767e211125..ec4402bf7a6 100644 --- a/test/test_transforms_v2.py +++ b/test/test_transforms_v2.py @@ -3355,6 +3355,9 @@ 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="CV-CUDA not available") + ), ], ) def test_functional(self, make_input): @@ -3370,9 +3373,16 @@ def test_functional(self, make_input): (F.elastic_mask, tv_tensors.Mask), (F.elastic_video, tv_tensors.Video), (F.elastic_keypoints, tv_tensors.KeyPoints), + pytest.param( + F._geometry._elastic_cvcuda, + "cvcuda.Tensor", + marks=pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="CV-CUDA not available"), + ), ], ) def test_functional_signature(self, kernel, input_type): + if input_type == "cvcuda.Tensor": + input_type = _import_cvcuda().Tensor check_functional_kernel_signature_match(F.elastic, kernel=kernel, input_type=input_type) @pytest.mark.parametrize( @@ -3385,6 +3395,9 @@ 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="CV-CUDA not available") + ), ], ) def test_displacement_error(self, make_input): @@ -3406,6 +3419,9 @@ def test_displacement_error(self, make_input): make_segmentation_mask, make_video, make_keypoints, + pytest.param( + make_image_cvcuda, marks=pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="CV-CUDA not available") + ), ], ) # ElasticTransform needs larger images to avoid the needed internal padding being larger than the actual image diff --git a/torchvision/transforms/v2/functional/_geometry.py b/torchvision/transforms/v2/functional/_geometry.py index 0e27218bc89..3897d395327 100644 --- a/torchvision/transforms/v2/functional/_geometry.py +++ b/torchvision/transforms/v2/functional/_geometry.py @@ -4,6 +4,7 @@ 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 @@ -2529,6 +2530,111 @@ def elastic_video( return elastic_image(video, displacement, interpolation=interpolation, fill=fill) +if CVCUDA_AVAILABLE: + _cvcuda_interp = { + InterpolationMode.BILINEAR: cvcuda.Interp.LINEAR, + "bilinear": cvcuda.Interp.LINEAR, + "linear": cvcuda.Interp.LINEAR, + 2: cvcuda.Interp.LINEAR, + InterpolationMode.BICUBIC: cvcuda.Interp.CUBIC, + "bicubic": cvcuda.Interp.CUBIC, + 3: cvcuda.Interp.CUBIC, + InterpolationMode.NEAREST: cvcuda.Interp.NEAREST, + "nearest": cvcuda.Interp.NEAREST, + 0: cvcuda.Interp.NEAREST, + InterpolationMode.BOX: cvcuda.Interp.BOX, + "box": cvcuda.Interp.BOX, + 4: cvcuda.Interp.BOX, + InterpolationMode.HAMMING: cvcuda.Interp.HAMMING, + "hamming": cvcuda.Interp.HAMMING, + 5: cvcuda.Interp.HAMMING, + InterpolationMode.LANCZOS: cvcuda.Interp.LANCZOS, + "lanczos": cvcuda.Interp.LANCZOS, + 1: cvcuda.Interp.LANCZOS, + } + + +def _elastic_cvcuda( + image: "cvcuda.Tensor", + displacement: torch.Tensor, + interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR, + fill: _FillTypeJIT = None, +) -> "cvcuda.Tensor": + if not isinstance(displacement, torch.Tensor): + raise TypeError("Argument displacement should be a Tensor") + + # Input image is NHWC format: (N, H, W, C) + batch_size, height, width, num_channels = image.shape + device = torch.device("cuda") + dtype = torch.float32 + + expected_shape = (1, height, width, 2) + if expected_shape != displacement.shape: + raise ValueError(f"Argument displacement shape should be {expected_shape}, but given {displacement.shape}") + + # cvcuda.remap only supports uint8 for 3-channel images, float32 for 1-channel + input_dtype = image.dtype + if num_channels == 3 and input_dtype != cvcuda.Type.U8: + raise ValueError(f"cvcuda.remap requires uint8 dtype for 3-channel images, but got {input_dtype}") + elif num_channels == 1 and input_dtype != cvcuda.Type.F32: + raise ValueError(f"cvcuda.remap requires float32 dtype for 1-channel images, but got {input_dtype}") + + # Build normalized grid: identity + displacement + # _create_identity_grid returns (1, H, W, 2) with values in [-1, 1] + identity_grid = _create_identity_grid((height, width), device=device, dtype=dtype) + grid = identity_grid.add_(displacement.to(dtype=dtype, device=device)) + + # Convert normalized grid [-1, 1] to absolute pixel coordinates [0, width-1], [0, height-1] + # grid[..., 0] is x (horizontal), grid[..., 1] is y (vertical) + map_x = (grid[..., 0] + 1) * (width - 1) / 2.0 + map_y = (grid[..., 1] + 1) * (height - 1) / 2.0 + + # Stack into (1, H, W, 2) map tensor + pixel_map = torch.stack([map_x, map_y], dim=-1) + + # Expand map for batch if needed + if batch_size > 1: + pixel_map = pixel_map.expand(batch_size, -1, -1, -1) + + # Create cvcuda map tensor (NHWC layout with 2 channels for x,y) + cv_map = cvcuda.as_tensor(pixel_map.contiguous(), "NHWC") + + # Resolve interpolation + src_interp = _cvcuda_interp.get(interpolation, cvcuda.Interp.LINEAR) + + # Resolve border mode and value + if fill is None: + border_mode = cvcuda.Border.CONSTANT + border_value = np.array([], dtype=np.float32) + elif isinstance(fill, (int, float)): + border_mode = cvcuda.Border.CONSTANT + border_value = np.array([fill], dtype=np.float32) + elif isinstance(fill, (list, tuple)): + border_mode = cvcuda.Border.CONSTANT + border_value = np.array(fill, dtype=np.float32) + else: + border_mode = cvcuda.Border.CONSTANT + border_value = np.array([], dtype=np.float32) + + # Call cvcuda.remap + output = cvcuda.remap( + image, + cv_map, + src_interp=src_interp, + map_interp=cvcuda.Interp.LINEAR, + map_type=cvcuda.Remap.ABSOLUTE, + align_corners=False, + border=border_mode, + border_value=border_value, + ) + + return output + + +if CVCUDA_AVAILABLE: + _elastic_cvcuda = _register_kernel_internal(elastic, cvcuda.Tensor)(_elastic_cvcuda) + + def center_crop(inpt: torch.Tensor, output_size: list[int]) -> torch.Tensor: """See :class:`~torchvision.transforms.v2.RandomCrop` for details.""" if torch.jit.is_scripting(): From 156893f279bfe7e01c0a9a7db9210f4a694948b9 Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Tue, 2 Dec 2025 12:49:28 -0800 Subject: [PATCH 10/15] update transformed types --- torchvision/transforms/v2/_geometry.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/torchvision/transforms/v2/_geometry.py b/torchvision/transforms/v2/_geometry.py index 96166e05e9a..6731134934d 100644 --- a/torchvision/transforms/v2/_geometry.py +++ b/torchvision/transforms/v2/_geometry.py @@ -29,6 +29,7 @@ is_pure_tensor, query_size, ) +from .functional._utils import is_cvcuda_tensor CVCUDA_AVAILABLE = _is_cvcuda_available() @@ -1045,6 +1046,8 @@ class ElasticTransform(Transform): _v1_transform_cls = _transforms.ElasticTransform + _transformed_types = Transform._transformed_types + (is_cvcuda_tensor,) + def __init__( self, alpha: Union[float, Sequence[float]] = 50.0, From c3621206211fb671bfd55e067504781671ec1990 Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Tue, 2 Dec 2025 12:52:58 -0800 Subject: [PATCH 11/15] simplify elastic cvcuda code more --- .../transforms/v2/functional/_geometry.py | 17 ++++++----------- 1 file changed, 6 insertions(+), 11 deletions(-) diff --git a/torchvision/transforms/v2/functional/_geometry.py b/torchvision/transforms/v2/functional/_geometry.py index 3897d395327..68ca6e6fbca 100644 --- a/torchvision/transforms/v2/functional/_geometry.py +++ b/torchvision/transforms/v2/functional/_geometry.py @@ -2563,7 +2563,6 @@ def _elastic_cvcuda( if not isinstance(displacement, torch.Tensor): raise TypeError("Argument displacement should be a Tensor") - # Input image is NHWC format: (N, H, W, C) batch_size, height, width, num_channels = image.shape device = torch.device("cuda") dtype = torch.float32 @@ -2579,6 +2578,10 @@ def _elastic_cvcuda( elif num_channels == 1 and input_dtype != cvcuda.Type.F32: raise ValueError(f"cvcuda.remap requires float32 dtype for 1-channel images, but got {input_dtype}") + interp = _cvcuda_interp.get(interpolation, cvcuda.Interp.LINEAR) + if interp is None: + raise ValueError(f"Invalid interpolation mode: {interpolation}") + # Build normalized grid: identity + displacement # _create_identity_grid returns (1, H, W, 2) with values in [-1, 1] identity_grid = _create_identity_grid((height, width), device=device, dtype=dtype) @@ -2599,28 +2602,20 @@ def _elastic_cvcuda( # Create cvcuda map tensor (NHWC layout with 2 channels for x,y) cv_map = cvcuda.as_tensor(pixel_map.contiguous(), "NHWC") - # Resolve interpolation - src_interp = _cvcuda_interp.get(interpolation, cvcuda.Interp.LINEAR) - - # Resolve border mode and value + border_mode = cvcuda.Border.CONSTANT if fill is None: - border_mode = cvcuda.Border.CONSTANT border_value = np.array([], dtype=np.float32) elif isinstance(fill, (int, float)): - border_mode = cvcuda.Border.CONSTANT border_value = np.array([fill], dtype=np.float32) elif isinstance(fill, (list, tuple)): - border_mode = cvcuda.Border.CONSTANT border_value = np.array(fill, dtype=np.float32) else: - border_mode = cvcuda.Border.CONSTANT border_value = np.array([], dtype=np.float32) - # Call cvcuda.remap output = cvcuda.remap( image, cv_map, - src_interp=src_interp, + src_interp=interp, map_interp=cvcuda.Interp.LINEAR, map_type=cvcuda.Remap.ABSOLUTE, align_corners=False, From be9183ad5822ef97ca5fcf10ab376a2b259285fc Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Thu, 4 Dec 2025 14:17:48 -0800 Subject: [PATCH 12/15] rebase to main standards --- test/test_transforms_v2.py | 7 ++- torchvision/transforms/v2/_geometry.py | 4 +- .../transforms/v2/functional/_augment.py | 11 +---- .../transforms/v2/functional/_color.py | 12 +---- .../transforms/v2/functional/_geometry.py | 35 +++----------- torchvision/transforms/v2/functional/_misc.py | 11 +---- .../transforms/v2/functional/_utils.py | 48 ++++++++++++++++++- 7 files changed, 62 insertions(+), 66 deletions(-) diff --git a/test/test_transforms_v2.py b/test/test_transforms_v2.py index ec4402bf7a6..8b3e68f51cb 100644 --- a/test/test_transforms_v2.py +++ b/test/test_transforms_v2.py @@ -25,7 +25,6 @@ assert_equal, cache, cpu_and_cuda, - cvcuda_to_pil_compatible_tensor, freeze_rng_state, ignore_jit_no_profile_information_warning, make_bounding_boxes, @@ -3374,14 +3373,14 @@ def test_functional(self, make_input): (F.elastic_video, tv_tensors.Video), (F.elastic_keypoints, tv_tensors.KeyPoints), pytest.param( - F._geometry._elastic_cvcuda, - "cvcuda.Tensor", + F._geometry._elastic_image_cvcuda, + None, marks=pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="CV-CUDA not available"), ), ], ) def test_functional_signature(self, kernel, input_type): - if input_type == "cvcuda.Tensor": + if kernel is F._geometry._elastic_image_cvcuda: input_type = _import_cvcuda().Tensor check_functional_kernel_signature_match(F.elastic, kernel=kernel, input_type=input_type) diff --git a/torchvision/transforms/v2/_geometry.py b/torchvision/transforms/v2/_geometry.py index 6731134934d..ada7b2a02ea 100644 --- a/torchvision/transforms/v2/_geometry.py +++ b/torchvision/transforms/v2/_geometry.py @@ -29,7 +29,6 @@ is_pure_tensor, query_size, ) -from .functional._utils import is_cvcuda_tensor CVCUDA_AVAILABLE = _is_cvcuda_available() @@ -1046,7 +1045,8 @@ class ElasticTransform(Transform): _v1_transform_cls = _transforms.ElasticTransform - _transformed_types = Transform._transformed_types + (is_cvcuda_tensor,) + if CVCUDA_AVAILABLE: + _transformed_types = Transform._transformed_types + (_is_cvcuda_tensor,) def __init__( self, diff --git a/torchvision/transforms/v2/functional/_augment.py b/torchvision/transforms/v2/functional/_augment.py index 7ce5bdc7b7e..a904d8d7cbd 100644 --- a/torchvision/transforms/v2/functional/_augment.py +++ b/torchvision/transforms/v2/functional/_augment.py @@ -1,5 +1,4 @@ import io -from typing import TYPE_CHECKING import PIL.Image @@ -9,15 +8,7 @@ from torchvision.transforms.functional import pil_to_tensor, to_pil_image from torchvision.utils import _log_api_usage_once -from ._utils import _get_kernel, _import_cvcuda, _is_cvcuda_available, _register_kernel_internal - - -CVCUDA_AVAILABLE = _is_cvcuda_available() - -if TYPE_CHECKING: - import cvcuda # type: ignore[import-not-found] -if CVCUDA_AVAILABLE: - cvcuda = _import_cvcuda() # noqa: F811 +from ._utils import _get_kernel, _register_kernel_internal def erase( diff --git a/torchvision/transforms/v2/functional/_color.py b/torchvision/transforms/v2/functional/_color.py index 5be9c62902a..be254c0d63a 100644 --- a/torchvision/transforms/v2/functional/_color.py +++ b/torchvision/transforms/v2/functional/_color.py @@ -1,5 +1,3 @@ -from typing import TYPE_CHECKING - import PIL.Image import torch from torch.nn.functional import conv2d @@ -11,15 +9,7 @@ from ._misc import _num_value_bits, to_dtype_image from ._type_conversion import pil_to_tensor, to_pil_image -from ._utils import _get_kernel, _import_cvcuda, _is_cvcuda_available, _register_kernel_internal - - -CVCUDA_AVAILABLE = _is_cvcuda_available() - -if TYPE_CHECKING: - import cvcuda # type: ignore[import-not-found] -if CVCUDA_AVAILABLE: - cvcuda = _import_cvcuda() # noqa: F811 +from ._utils import _get_kernel, _register_kernel_internal def rgb_to_grayscale(inpt: torch.Tensor, num_output_channels: int = 1) -> torch.Tensor: diff --git a/torchvision/transforms/v2/functional/_geometry.py b/torchvision/transforms/v2/functional/_geometry.py index 68ca6e6fbca..aa63e69934f 100644 --- a/torchvision/transforms/v2/functional/_geometry.py +++ b/torchvision/transforms/v2/functional/_geometry.py @@ -29,6 +29,7 @@ from ._utils import ( _FillTypeJIT, + _get_cvcuda_interp, _get_kernel, _import_cvcuda, _is_cvcuda_available, @@ -2530,36 +2531,14 @@ def elastic_video( return elastic_image(video, displacement, interpolation=interpolation, fill=fill) -if CVCUDA_AVAILABLE: - _cvcuda_interp = { - InterpolationMode.BILINEAR: cvcuda.Interp.LINEAR, - "bilinear": cvcuda.Interp.LINEAR, - "linear": cvcuda.Interp.LINEAR, - 2: cvcuda.Interp.LINEAR, - InterpolationMode.BICUBIC: cvcuda.Interp.CUBIC, - "bicubic": cvcuda.Interp.CUBIC, - 3: cvcuda.Interp.CUBIC, - InterpolationMode.NEAREST: cvcuda.Interp.NEAREST, - "nearest": cvcuda.Interp.NEAREST, - 0: cvcuda.Interp.NEAREST, - InterpolationMode.BOX: cvcuda.Interp.BOX, - "box": cvcuda.Interp.BOX, - 4: cvcuda.Interp.BOX, - InterpolationMode.HAMMING: cvcuda.Interp.HAMMING, - "hamming": cvcuda.Interp.HAMMING, - 5: cvcuda.Interp.HAMMING, - InterpolationMode.LANCZOS: cvcuda.Interp.LANCZOS, - "lanczos": cvcuda.Interp.LANCZOS, - 1: cvcuda.Interp.LANCZOS, - } - - -def _elastic_cvcuda( +def _elastic_image_cvcuda( image: "cvcuda.Tensor", displacement: torch.Tensor, interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR, fill: _FillTypeJIT = None, ) -> "cvcuda.Tensor": + cvcuda = _import_cvcuda() + if not isinstance(displacement, torch.Tensor): raise TypeError("Argument displacement should be a Tensor") @@ -2578,9 +2557,7 @@ def _elastic_cvcuda( elif num_channels == 1 and input_dtype != cvcuda.Type.F32: raise ValueError(f"cvcuda.remap requires float32 dtype for 1-channel images, but got {input_dtype}") - interp = _cvcuda_interp.get(interpolation, cvcuda.Interp.LINEAR) - if interp is None: - raise ValueError(f"Invalid interpolation mode: {interpolation}") + interp = _get_cvcuda_interp(interpolation) # Build normalized grid: identity + displacement # _create_identity_grid returns (1, H, W, 2) with values in [-1, 1] @@ -2627,7 +2604,7 @@ def _elastic_cvcuda( if CVCUDA_AVAILABLE: - _elastic_cvcuda = _register_kernel_internal(elastic, cvcuda.Tensor)(_elastic_cvcuda) + _register_kernel_internal(elastic, _import_cvcuda().Tensor)(_elastic_image_cvcuda) def center_crop(inpt: torch.Tensor, output_size: list[int]) -> torch.Tensor: diff --git a/torchvision/transforms/v2/functional/_misc.py b/torchvision/transforms/v2/functional/_misc.py index 0fa05a2113c..daf263df046 100644 --- a/torchvision/transforms/v2/functional/_misc.py +++ b/torchvision/transforms/v2/functional/_misc.py @@ -1,5 +1,5 @@ import math -from typing import Optional, TYPE_CHECKING +from typing import Optional import PIL.Image import torch @@ -13,14 +13,7 @@ from ._meta import _convert_bounding_box_format -from ._utils import _get_kernel, _import_cvcuda, _is_cvcuda_available, _register_kernel_internal, is_pure_tensor - -CVCUDA_AVAILABLE = _is_cvcuda_available() - -if TYPE_CHECKING: - import cvcuda # type: ignore[import-not-found] -if CVCUDA_AVAILABLE: - cvcuda = _import_cvcuda() # noqa: F811 +from ._utils import _get_kernel, _register_kernel_internal, is_pure_tensor def normalize( diff --git a/torchvision/transforms/v2/functional/_utils.py b/torchvision/transforms/v2/functional/_utils.py index 11480b30ef9..963f50e08cb 100644 --- a/torchvision/transforms/v2/functional/_utils.py +++ b/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 +181,45 @@ 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 _populate_interpolation_mode_to_cvcuda_interp(): + cvcuda = _import_cvcuda() + + global _interpolation_mode_to_cvcuda_interp + + _interpolation_mode_to_cvcuda_interp = { + InterpolationMode.BILINEAR: cvcuda.Interp.LINEAR, + "bilinear": cvcuda.Interp.LINEAR, + "linear": cvcuda.Interp.LINEAR, + 2: cvcuda.Interp.LINEAR, + InterpolationMode.BICUBIC: cvcuda.Interp.CUBIC, + "bicubic": cvcuda.Interp.CUBIC, + 3: cvcuda.Interp.CUBIC, + InterpolationMode.NEAREST: cvcuda.Interp.NEAREST, + "nearest": cvcuda.Interp.NEAREST, + 0: cvcuda.Interp.NEAREST, + InterpolationMode.BOX: cvcuda.Interp.BOX, + "box": cvcuda.Interp.BOX, + 4: cvcuda.Interp.BOX, + InterpolationMode.HAMMING: cvcuda.Interp.HAMMING, + "hamming": cvcuda.Interp.HAMMING, + 5: cvcuda.Interp.HAMMING, + InterpolationMode.LANCZOS: cvcuda.Interp.LANCZOS, + "lanczos": cvcuda.Interp.LANCZOS, + 1: cvcuda.Interp.LANCZOS, + } + + +def _get_cvcuda_interp(interpolation: InterpolationMode | str | int) -> "cvcuda.Interp": + if len(_interpolation_mode_to_cvcuda_interp) == 0: + _populate_interpolation_mode_to_cvcuda_interp() + + 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 From ff682531821b098fcf8652290e1e0e2879a9b20f Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Thu, 4 Dec 2025 18:07:39 -0800 Subject: [PATCH 13/15] update to resize interps --- torchvision/transforms/v2/functional/_utils.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/torchvision/transforms/v2/functional/_utils.py b/torchvision/transforms/v2/functional/_utils.py index 963f50e08cb..4111416df79 100644 --- a/torchvision/transforms/v2/functional/_utils.py +++ b/torchvision/transforms/v2/functional/_utils.py @@ -191,6 +191,8 @@ def _populate_interpolation_mode_to_cvcuda_interp(): global _interpolation_mode_to_cvcuda_interp + # CV-CUDA's NEAREST matches PyTorch's 'nearest-exact' (PIL-style) + # not PyTorch's 'nearest' (OpenCV-style). _interpolation_mode_to_cvcuda_interp = { InterpolationMode.BILINEAR: cvcuda.Interp.LINEAR, "bilinear": cvcuda.Interp.LINEAR, @@ -202,6 +204,8 @@ def _populate_interpolation_mode_to_cvcuda_interp(): InterpolationMode.NEAREST: cvcuda.Interp.NEAREST, "nearest": cvcuda.Interp.NEAREST, 0: cvcuda.Interp.NEAREST, + InterpolationMode.NEAREST_EXACT: cvcuda.Interp.NEAREST, + "nearest-exact": cvcuda.Interp.NEAREST, InterpolationMode.BOX: cvcuda.Interp.BOX, "box": cvcuda.Interp.BOX, 4: cvcuda.Interp.BOX, From 62eeddb19d3013759a6a7551a46c51a8975d676b Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Thu, 4 Dec 2025 18:16:30 -0800 Subject: [PATCH 14/15] refactor interp setup --- .../transforms/v2/functional/_utils.py | 54 ++++++++----------- 1 file changed, 21 insertions(+), 33 deletions(-) diff --git a/torchvision/transforms/v2/functional/_utils.py b/torchvision/transforms/v2/functional/_utils.py index 4111416df79..a1742ba149f 100644 --- a/torchvision/transforms/v2/functional/_utils.py +++ b/torchvision/transforms/v2/functional/_utils.py @@ -186,41 +186,29 @@ def _is_cvcuda_tensor(inpt: Any) -> bool: _interpolation_mode_to_cvcuda_interp: dict[InterpolationMode | str | int, "cvcuda.Interp"] = {} -def _populate_interpolation_mode_to_cvcuda_interp(): - cvcuda = _import_cvcuda() - - global _interpolation_mode_to_cvcuda_interp - - # CV-CUDA's NEAREST matches PyTorch's 'nearest-exact' (PIL-style) - # not PyTorch's 'nearest' (OpenCV-style). - _interpolation_mode_to_cvcuda_interp = { - InterpolationMode.BILINEAR: cvcuda.Interp.LINEAR, - "bilinear": cvcuda.Interp.LINEAR, - "linear": cvcuda.Interp.LINEAR, - 2: cvcuda.Interp.LINEAR, - InterpolationMode.BICUBIC: cvcuda.Interp.CUBIC, - "bicubic": cvcuda.Interp.CUBIC, - 3: cvcuda.Interp.CUBIC, - InterpolationMode.NEAREST: cvcuda.Interp.NEAREST, - "nearest": cvcuda.Interp.NEAREST, - 0: cvcuda.Interp.NEAREST, - InterpolationMode.NEAREST_EXACT: cvcuda.Interp.NEAREST, - "nearest-exact": cvcuda.Interp.NEAREST, - InterpolationMode.BOX: cvcuda.Interp.BOX, - "box": cvcuda.Interp.BOX, - 4: cvcuda.Interp.BOX, - InterpolationMode.HAMMING: cvcuda.Interp.HAMMING, - "hamming": cvcuda.Interp.HAMMING, - 5: cvcuda.Interp.HAMMING, - InterpolationMode.LANCZOS: cvcuda.Interp.LANCZOS, - "lanczos": cvcuda.Interp.LANCZOS, - 1: cvcuda.Interp.LANCZOS, - } - - def _get_cvcuda_interp(interpolation: InterpolationMode | str | int) -> "cvcuda.Interp": if len(_interpolation_mode_to_cvcuda_interp) == 0: - _populate_interpolation_mode_to_cvcuda_interp() + 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: From 0f8ff252a2b8ac5adce36ec3b82fde7df0a45a5a Mon Sep 17 00:00:00 2001 From: Justin Davis Date: Fri, 12 Dec 2025 11:42:56 -0800 Subject: [PATCH 15/15] add additional test on correctness for cvcuda compared to torchvision, minimize diff --- test/test_transforms_v2.py | 31 +++++++++++++++++++ torchvision/transforms/v2/_geometry.py | 3 +- torchvision/transforms/v2/_utils.py | 2 +- torchvision/transforms/v2/functional/_meta.py | 22 +------------ .../transforms/v2/functional/_utils.py | 8 +++++ 5 files changed, 42 insertions(+), 24 deletions(-) diff --git a/test/test_transforms_v2.py b/test/test_transforms_v2.py index 8b3e68f51cb..f34e6a730d8 100644 --- a/test/test_transforms_v2.py +++ b/test/test_transforms_v2.py @@ -3438,6 +3438,37 @@ def test_transform(self, make_input, size, device): check_v1_compatibility=check_v1_compatibility, ) + @pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="CV-CUDA not available") + @needs_cuda + @pytest.mark.parametrize( + "interpolation", [transforms.InterpolationMode.NEAREST, transforms.InterpolationMode.BILINEAR] + ) + def test_image_cvcuda_correctness(self, interpolation): + image = make_image_cvcuda(dtype=torch.uint8) + displacement = self._make_displacement(image) + + result = F._geometry._elastic_image_cvcuda(image, displacement=displacement, interpolation=interpolation) + result = F.cvcuda_to_tensor(result) + + expected = F._geometry.elastic_image( + F.cvcuda_to_tensor(image), displacement=displacement, interpolation=interpolation + ) + + # mainly for checking properties (outside pixel values) are correct + # see note below on pixel-value differences + assert_close(result, expected, atol=get_max_value(torch.uint8), rtol=0) + + # visually, the results are identical, however the underlying computations are different + # we can define an mae_threshold based on the interpolation mode + # the primary difference is along the borders where pixels appear to be shifted in location + # by up to 1, causing potentially up to a diff of 255 on a single pixel + # this could be because one has fill of 0 and CV-CUDA is shifted and has value with some color + # thresholds decrease as image size gets larger + # (640, 480) input, has 20.0, 13.0 respectively to pass + mae = (expected.float() - result.float()).abs().mean() + mae_threshold = 30.0 if interpolation is transforms.InterpolationMode.NEAREST else 20.0 + assert mae < mae_threshold, f"MAE {mae} exceeds threshold" + class TestToPureTensor: def test_correctness(self): diff --git a/torchvision/transforms/v2/_geometry.py b/torchvision/transforms/v2/_geometry.py index ada7b2a02ea..7bc84112676 100644 --- a/torchvision/transforms/v2/_geometry.py +++ b/torchvision/transforms/v2/_geometry.py @@ -1045,8 +1045,7 @@ class ElasticTransform(Transform): _v1_transform_cls = _transforms.ElasticTransform - if CVCUDA_AVAILABLE: - _transformed_types = Transform._transformed_types + (_is_cvcuda_tensor,) + _transformed_types = Transform._transformed_types + (_is_cvcuda_tensor,) def __init__( self, diff --git a/torchvision/transforms/v2/_utils.py b/torchvision/transforms/v2/_utils.py index e803aa49c60..7274abaa861 100644 --- a/torchvision/transforms/v2/_utils.py +++ b/torchvision/transforms/v2/_utils.py @@ -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, _is_cvcuda_tensor)) + if check_type(inpt, (is_pure_tensor, tv_tensors.Image, PIL.Image.Image, tv_tensors.Video)) } if not chws: raise TypeError("No image or video was found in the sample") diff --git a/torchvision/transforms/v2/functional/_meta.py b/torchvision/transforms/v2/functional/_meta.py index af03ad018d4..6b8f19f12f4 100644 --- a/torchvision/transforms/v2/functional/_meta.py +++ b/torchvision/transforms/v2/functional/_meta.py @@ -51,16 +51,6 @@ 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) @@ -97,16 +87,6 @@ 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) @@ -145,7 +125,7 @@ def get_size_image_cvcuda(image: "cvcuda.Tensor") -> list[int]: if CVCUDA_AVAILABLE: - _register_kernel_internal(get_size, _import_cvcuda().Tensor)(get_size_image_cvcuda) + _get_size_image_cvcuda = _register_kernel_internal(get_size, cvcuda.Tensor)(get_size_image_cvcuda) @_register_kernel_internal(get_size, tv_tensors.Video, tv_tensor_wrapper=False) diff --git a/torchvision/transforms/v2/functional/_utils.py b/torchvision/transforms/v2/functional/_utils.py index a1742ba149f..b924bb16d38 100644 --- a/torchvision/transforms/v2/functional/_utils.py +++ b/torchvision/transforms/v2/functional/_utils.py @@ -187,6 +187,14 @@ def _is_cvcuda_tensor(inpt: Any) -> bool: def _get_cvcuda_interp(interpolation: InterpolationMode | str | int) -> "cvcuda.Interp": + """ + Get the CV-CUDA interpolation mode for a given interpolation mode. + + CV-CUDA has the two following differences (evaluated in tests) comapred to TorchVision/PIL: + 1. CV-CUDA does not have a match for NEAREST, its Interp.NEAREST is actually NEAREST_EXACT + Since we need to do interpolation, we will map NEAREST to Interp.NEAREST (which is NEAREST_EXACT) + 2. BICUBIC interpolation method is different compared to TorchVision/PIL, algorithmic difference + """ if len(_interpolation_mode_to_cvcuda_interp) == 0: cvcuda = _import_cvcuda() _interpolation_mode_to_cvcuda_interp[InterpolationMode.NEAREST] = cvcuda.Interp.NEAREST