masked_scatter

→Tensor

masked_scatter(input: Tensor, mask: Tensor, source: Tensor)

source edit

Fill the positions where mask holds with source's leading elements.

input and mask broadcast together; the k-th position the mask selects, in row-major order, takes source.reshape(-1)[k], and every other position keeps input's value. Differentiable through input (where the mask is False) and source (its elements that were written).

Parameters

inputTensor

Destination values; not mutated.

maskTensor

Bool tensor broadcasting against input.

sourceTensor

Values to write, of input's dtype and any shape, with at least as many elements as the mask selects.

Returns

Tensor

A new tensor of the broadcast shape of input and mask, in input's dtype — never input itself, even when the mask selects nothing.

Raises

lucid._C.engine.DtypeMismatch

mask is not bool, or source's dtype is not input's.

lucid._C.engine.ShapeMismatch

source has fewer elements than the mask selects.

IndexError

input and mask do not broadcast together.

Examples

>>> import lucid
>>> mask = lucid.tensor([True, False])
>>> lucid.masked_scatter(lucid.zeros(2, 2), mask, lucid.tensor([1.0, 2.0, 3.0]))
tensor([[1., 0.], [2., 0.]])