albumentations.core.tensor

CPU Tensor input contract and annotation bridge shared by Compose.

Members

functionvalidate_tensor_input
functiontensor_to_numpy_annotation
functiontensor_to_numpy_spatial
functionnumpy_to_tensor_spatial
functionnumpy_to_tensor_annotation

validate_tensor_inputfunction

validate_tensor_input(
    value: torch.Tensor,
    data_name: str,
    canonical_name: str
)

Validate a CPU Tensor against target-specific Compose shape, dtype, device, layout, and autograd boundary rules before it reaches a transform helper. The CPU stage accepts explicit-channel image targets, spatial mask targets, and float32 annotation matrices. It preserves non-contiguous strides because each accepted capability later decides whether a contiguous copy is justified by its full-path benchmark.

Parameters

Name	Type	Default	Description
value	torch.Tensor	-	-
data_name	str	-	-
canonical_name	str	-	-

tensor_to_numpy_annotationfunction

tensor_to_numpy_annotation(
    value: torch.Tensor,
    target: str
)

Return a NumPy view of a validated Tensor bbox or keypoint matrix through the shared annotation bridge used by the existing geometry processors.

Parameters

Name	Type	Default	Description
value	torch.Tensor	-	-
target	str	-	-

tensor_to_numpy_spatialfunction

tensor_to_numpy_spatial(
    value: torch.Tensor,
    target: str
)

Convert a validated Tensor target to the channel-last NumPy layout that existing Compose preprocessing and transforms expect. The bridge owns every layout conversion. Transform helpers only receive their established NumPy layout and never decide whether to convert a caller's Tensor themselves.

Parameters

Name	Type	Default	Description
value	torch.Tensor	-	-
target	str	-	-

numpy_to_tensor_spatialfunction

numpy_to_tensor_spatial(
    value: NDArray[np.generic],
    target: str
)

Convert a NumPy result from Compose back to the caller-facing Tensor layout, copying only when negative strides require it. NumPy transforms may return a negative-stride view, such as after a reflection. PyTorch cannot share that storage, so materialize only that incompatible case before returning the Tensor result.

Parameters

Name	Type	Default	Description
value	NDArray[np.generic]	-	-
target	str	-	-

numpy_to_tensor_annotationfunction

numpy_to_tensor_annotation(
    value: NDArray[np.generic],
    target: str
)

Return a Tensor bbox or keypoint matrix from a processor result, materializing only negative-stride NumPy storage that PyTorch cannot safely share.

Parameters

Name	Type	Default	Description
value	NDArray[np.generic]	-	-
target	str	-	-