review the transform.backend variable

monai/transforms/croppad/array.py

@@ -363,7 +363,7 @@ def compute_pad_width(self, spatial_shape: Sequence[int]) -> List[Tuple[int, int
 
 class Crop(InvertibleTransform):
     """
-    Perform crop operation on the input image.
+    Perform crop operations on the input image.
 
     """
 
@@ -378,7 +378,7 @@ def compute_slices(
         roi_slices: Optional[Sequence[slice]] = None,
     ):
         """
-        Compute the crop slices based on specified `center & size` or `start & end`.
+        Compute the crop slices based on specified `center & size` or `start & end` or `slices`.
 
         Args:
             roi_center: voxel coordinates for center of the crop ROI.
@@ -471,7 +471,7 @@ class SpatialCrop(Crop):
     It can support to crop ND spatial (channel-first) data.
 
     The cropped region can be parameterised in various ways:
-        - a list of slices for each spatial dimension (allows for use of -ve indexing and `None`)
+        - a list of slices for each spatial dimension (allows for use of negative indexing and `None`)
         - a spatial center and size
         - the start and end coordinates of the ROI
     """

monai/transforms/meta_utility/dictionary.py

@@ -45,7 +45,7 @@ class FromMetaTensord(MapTransform, InvertibleTransform):
     have the form `{"a": torch.Tensor, "a_meta_dict": dict, "a_transforms": list, "b": ...}`.
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.TORCH, TransformBackends.NUMPY, TransformBackends.CUPY]
 
     def __init__(
         self, keys: KeysCollection, data_type: Union[Sequence[str], str] = "tensor", allow_missing_keys: bool = False
@@ -92,7 +92,7 @@ class ToMetaTensord(MapTransform, InvertibleTransform):
     have the form `{"a": MetaTensor, "b": MetaTensor}`.
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.TORCH, TransformBackends.NUMPY, TransformBackends.CUPY]
 
     def __call__(self, data: Mapping[Hashable, NdarrayOrTensor]) -> Dict[Hashable, NdarrayOrTensor]:
         d = dict(data)

monai/transforms/post/array.py

@@ -265,7 +265,7 @@ class KeepLargestConnectedComponent(Transform):
 
     """
 
-    backend = [TransformBackends.NUMPY]
+    backend = [TransformBackends.NUMPY, TransformBackends.CUPY]
 
     def __init__(
         self,
@@ -354,6 +354,8 @@ class RemoveSmallObjects(Transform):
             If false, the overall size islands made from all non-background voxels will be used.
     """
 
+    backend = [TransformBackends.NUMPY]
+
     def __init__(self, min_size: int = 64, connectivity: int = 1, independent_channels: bool = True) -> None:
         self.min_size = min_size
         self.connectivity = connectivity
@@ -703,7 +705,7 @@ class ProbNMS(Transform):
 
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.NUMPY]
 
     def __init__(
         self,
@@ -765,6 +767,8 @@ class Invert(Transform):
     Utility transform to automatically invert the previously applied transforms.
     """
 
+    backend = [TransformBackends.TORCH]
+
     def __init__(
         self,
         transform: Optional[InvertibleTransform] = None,

monai/transforms/post/dictionary.py

@@ -254,6 +254,8 @@ class RemoveSmallObjectsd(MapTransform):
             If false, the overall size islands made from all non-background voxels will be used.
     """
 
+    backend = RemoveSmallObjects.backend
+
     def __init__(
         self,
         keys: KeysCollection,
@@ -785,6 +787,8 @@ class SobelGradientsd(MapTransform):
 
     """
 
+    backend = SobelGradients.backend
+
     def __init__(
         self,
         keys: KeysCollection,

monai/transforms/smooth_field/array.py

@@ -52,6 +52,8 @@ class SmoothField(Randomizable):
         device: Pytorch device to define field on
     """
 
+    backend = [TransformBackends.TORCH]
+
     def __init__(
         self,
         rand_size: Sequence[int],
@@ -160,7 +162,7 @@ class RandSmoothFieldAdjustContrast(RandomizableTransform):
         device: Pytorch device to define field on
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.TORCH]
 
     def __init__(
         self,
@@ -261,7 +263,7 @@ class RandSmoothFieldAdjustIntensity(RandomizableTransform):
         device: Pytorch device to define field on
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.TORCH]
 
     def __init__(
         self,

monai/transforms/smooth_field/dictionary.py

@@ -25,7 +25,6 @@
 )
 from monai.transforms.transform import MapTransform, RandomizableTransform
 from monai.utils import GridSampleMode, GridSamplePadMode, InterpolateMode, convert_to_tensor, ensure_tuple_rep
-from monai.utils.enums import TransformBackends
 
 __all__ = [
     "RandSmoothFieldAdjustContrastd",
@@ -60,7 +59,7 @@ class RandSmoothFieldAdjustContrastd(RandomizableTransform, MapTransform):
         device: Pytorch device to define field on
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = RandSmoothFieldAdjustContrast.backend
 
     def __init__(
         self,
@@ -138,7 +137,7 @@ class RandSmoothFieldAdjustIntensityd(RandomizableTransform, MapTransform):
         device: Pytorch device to define field on
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = RandSmoothFieldAdjustIntensity.backend
 
     def __init__(
         self,
@@ -219,7 +218,7 @@ class RandSmoothDeformd(RandomizableTransform, MapTransform):
         device: Pytorch device to define field on
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = RandSmoothDeform.backend
 
     def __init__(
         self,

monai/transforms/spatial/array.py

@@ -115,7 +115,7 @@ class SpatialResample(InvertibleTransform):
     by ``xform = linalg.solve(src_affine, dst_affine)``, and call ``monai.transforms.Affine`` with ``xform``.
     """
 
-    backend = [TransformBackends.TORCH]
+    backend = [TransformBackends.TORCH, TransformBackends.NUMPY, TransformBackends.CUPY]
 
     def __init__(
         self,

monai/transforms/transform.py

@@ -215,9 +215,13 @@ class Transform(ABC):
         :py:class:`monai.transforms.Compose`
     """
 
-    # Transforms should add data types to this list if they are capable of performing a transform without
-    # modifying the input type. For example, ["torch.Tensor", "np.ndarray"] means that no copies of the data
-    # are required if the input is either `torch.Tensor` or `np.ndarray`.
+    # Transforms should add `monai.transforms.utils.TransformBackends` to this list if they are performing
+    # the data processing using the corresponding backend APIs.
+    # Most of MONAI transform's inputs and outputs will be converted into torch.Tensor or monai.data.MetaTensor.
+    # This variable provides information about whether the input will be converted
+    # to other data types during the transformation. Note that not all `dtype` (such as float32, uint8) are supported
+    # by all the data types, the `dtype` during the conversion is determined automatically by each transform,
+    # please refer to the transform's docstring.
     backend: List[TransformBackends] = []
 
     @abstractmethod

monai/transforms/utility/array.py

@@ -269,7 +269,7 @@ class RepeatChannel(Transform):
         repeats: the number of repetitions for each element.
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.TORCH]
 
     def __init__(self, repeats: int) -> None:
         if repeats <= 0:
@@ -424,7 +424,7 @@ class ToTensor(Transform):
 
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.TORCH]
 
     def __init__(
         self,
@@ -519,7 +519,7 @@ class ToNumpy(Transform):
 
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.NUMPY]
 
     def __init__(self, dtype: DtypeLike = None, wrap_sequence: bool = True) -> None:
         super().__init__()
@@ -546,7 +546,7 @@ class ToCupy(Transform):
 
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.CUPY]
 
     def __init__(self, dtype: Optional[np.dtype] = None, wrap_sequence: bool = True) -> None:
         super().__init__()
@@ -565,7 +565,7 @@ class ToPIL(Transform):
     Converts the input image (in the form of NumPy array or PyTorch Tensor) to PIL image
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.NUMPY]
 
     def __call__(self, img):
         """
@@ -583,7 +583,7 @@ class Transpose(Transform):
     Transposes the input image based on the given `indices` dimension ordering.
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.TORCH]
 
     def __init__(self, indices: Optional[Sequence[int]]) -> None:
         self.indices = None if indices is None else tuple(indices)
@@ -1090,7 +1090,7 @@ class AddExtremePointsChannel(Randomizable, Transform):
         ValueError: When label image is not single channel.
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.TORCH]
 
     def __init__(self, background: int = 0, pert: float = 0.0) -> None:
         self._background = background
@@ -1422,7 +1422,7 @@ class AddCoordinateChannels(Transform):
 
     """
 
-    backend = [TransformBackends.TORCH, TransformBackends.NUMPY]
+    backend = [TransformBackends.NUMPY]
 
     @deprecated_arg(
         name="spatial_channels", new_name="spatial_dims", since="0.8", msg_suffix="please use `spatial_dims` instead."

monai/transforms/utility/dictionary.py

@@ -377,6 +377,9 @@ def __call__(self, data: Mapping[Hashable, NdarrayOrTensor]) -> Dict[Hashable, N
 
 
 class SplitDimd(MapTransform):
+
+    backend = SplitDim.backend
+
     def __init__(
         self,
         keys: KeysCollection,

monai/utils/enums.py

@@ -381,11 +381,16 @@ def transforms(key: Optional[str] = None):
 
 class TransformBackends(StrEnum):
     """
-    Transform backends.
+    Transform backends. Most of `monai.transforms` components first converts the input data into ``torch.Tensor`` or
+    ``monai.data.MetaTensor``. Internally, some transforms are made by converting the data into ``numpy.array`` or
+    ``cupy.array`` and use the underlying transform backend API to achieve the actual output array and
+    converting back to ``Tensor``/``MetaTensor``. Transforms with more than one backend indicate the that they may
+    convert the input data types to accomodate the underlying API.
     """
 
     TORCH = "torch"
     NUMPY = "numpy"
+    CUPY = "cupy"
 
 
 class JITMetadataKeys(StrEnum):