lucid.empty_like¶

lucid.empty_like(a: Tensor | list | ndarray | array, /, dtype: type[bool | int | float | complex] | Numeric | None = None, requires_grad: bool = False, keep_grad: bool = False, device: Literal['cpu', 'gpu'] | None = None) → Tensor¶

The empty_like function generates a tensor with the same shape as the input tensor, with uninitialized values in its memory. The data type and gradient tracking behavior can be specified independently of the input tensor.

Function Signature¶

def empty_like(
    a: Tensor | _ArrayLike,
    dtype: type | None = None,
    requires_grad: bool = False,
    keep_grad: bool = False,
    device: _DeviceType | None = None,
) -> Tensor

Parameters¶

a (Tensor or _ArrayLike): The input tensor whose shape is used to define the shape of the output tensor.
dtype (type, optional): The data type of the uninitialized tensor. If None, the data type of a is used. Defaults to None.
requires_grad (bool, optional): If set to True, the resulting tensor will track gradients for automatic differentiation. Defaults to False.
keep_grad (bool, optional): Determines whether gradient history should persist across multiple operations. Defaults to False.

Returns¶

Tensor: A tensor with the same shape as a, containing uninitialized values.

Example¶

>>> import lucid
>>> a = lucid.Tensor([[1, 2, 3], [4, 5, 6]])
>>> x = lucid.empty_like(a)
>>> print(x)
Tensor([[ 6.949e-310,  6.949e-310,  0.000e+000],
        [ 0.000e+000,  0.000e+000,  0.000e+000]], grad=None)

The memory is uninitialized, so values are arbitrary. To track gradients, set requires_grad=True:

>>> y = lucid.empty_like(a, requires_grad=True)
>>> print(y.requires_grad)
True

You can also specify the data type of the new tensor:

>>> z = lucid.empty_like(a, dtype=float)
>>> print(z.dtype)
float

Note

The empty_like function does not clear the memory where the tensor is allocated, so the values in the tensor are arbitrary. If you need a tensor with zeros, consider using lucid.zeros_like instead.