as_tensor

→Tensor

as_tensor(data: object, dtype: DTypeLike = None, device: DeviceLike = None)

source edit

Convert data to a tensor, avoiding a copy when the source already matches.

Unlike tensor, as_tensor is "best-effort no-copy":

If data is already a Tensor with the requested dtype and device, it is returned unchanged.
If data is a Tensor whose dtype or device differs, it is converted as Tensor.to converts it, and the result stays connected to data's autograd graph — unlike tensor, which copies into a detached leaf.
Anything else — Python data, NumPy arrays — goes through tensor, which copies.

Parameters

dataobject

Source data — Python scalar / list, NumPy array, or Tensor.

dtypedtype | str | None

Target element type. None preserves the source dtype.

devicedevice | str | None

Target device. When the source already lives on a different device, a copy across the device boundary is performed.

Returns

Tensor

data itself, a converted copy of it, or a freshly-constructed Lucid tensor.

Notes

as_tensor is the right choice in code that may be handed either a Tensor or raw data (e.g. a collate function): a Tensor that already has the requested dtype and device passes through with no allocation at all. For semantic clarity in library code that should never share storage, use tensor.

Examples

>>> import lucid
>>> x = lucid.tensor([1.0, 2.0, 3.0])
>>> lucid.as_tensor(x) is x              # already a Tensor, returned as-is
True
>>> lucid.as_tensor(x, dtype=lucid.float64).dtype
lucid.float64
>>> w = lucid.tensor([1.0, 2.0], requires_grad=True)
>>> lucid.as_tensor(w, dtype=lucid.float64).sum().backward()
>>> w.grad                               # the conversion stays in the graph
tensor([1., 1.])