masked_fill

→Tensor

masked_fill(input: Tensor, mask: Tensor, value: _float)

source edit

Implementing kernel

C++masked_fill_opfree fn

Return input with positions where mask is True set to value.

The mask must be broadcast-compatible with input. Out-of-place: a fresh tensor is returned; input is not modified.

Parameters

inputTensor

Source tensor.

maskTensor

Boolean tensor broadcastable to input.

valuefloat

Scalar fill value.

Returns

Tensor

Tensor of the same shape and dtype as input.

Notes

Often used to apply attention masks:

\text{out}_i \;=\; \begin{cases} \text{value} & \text{if } \text{mask}_i \\ \text{input}_i & \text{otherwise} . \end{cases}

Examples

>>> import lucid
>>> x = lucid.tensor([1.0, 2.0, 3.0])
>>> m = lucid.tensor([True, False, True])
>>> lucid.masked_fill(x, m, 0.0)
Tensor([0., 2., 0.])