nansum

→Tensor

nansum(x: Tensor, dim: int | Sequence[int] | None = ..., keepdim: bool = ...)

Sum the tensor, treating NaN entries as zero.

A NaN-safe variant of lucid.sum. NaN values are replaced by 0 before the reduction, so they neither contaminate the result nor contribute to it.

xTensor

Input tensor (any floating-point dtype).

dimint | Sequence[int] | None

Axis or axes along which to sum. None (default) reduces over the entire tensor.

keepdimbool

If True, retains the reduced dimensions with size 1. Defaults to False.

Tensor

Reduced tensor.

With $\mathcal{S}$ the set of indices being reduced and $\mathbb{1}_{x_i \text{ is NaN}}$ an indicator,

\text{nansum}(x) = \sum_{i \in \mathcal{S}} (1 - \mathbb{1}_{x_i \text{ is NaN}}) \cdot x_i.

The gradient at NaN positions is zero — those entries contribute nothing to the forward sum, so they cannot influence it through perturbation.

>>> import lucid
>>> import math
>>> x = lucid.tensor([1.0, math.nan, 3.0])
>>> lucid.nansum(x)
Tensor(4.)