nanmean

→Tensor

nanmean(input: Tensor, dim: _int | Sequence[_int] | None = ..., keepdim: _bool = ..., dtype: DTypeLike | None = ...)

Mean of the tensor, ignoring NaN entries.

A NaN-safe variant of lucid.mean. Both the numerator (sum) and the denominator (element count) are computed only over the non-NaN entries.

inputTensor

Input tensor (any floating-point dtype).

dimint | Sequence[int] | None

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

keepdimbool

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

dtypeDTypeLike | None

Target floating-point or complex dtype for the reduction. When given, input is cast to dtype before averaging.

Tensor

Reduced tensor.

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

\text{nanmean}(x) = \frac{\sum_{i \in \mathcal{S}} \mathbb{1}_i \cdot x_i} {\sum_{i \in \mathcal{S}} \mathbb{1}_i}.

A slice that is entirely NaN produces 0 / 0 = NaN; this is the standard convention. The gradient at NaN positions is zero.

>>> import lucid
>>> import math
>>> x = lucid.tensor([1.0, math.nan, 3.0])
>>> lucid.nanmean(x)
tensor(2.)