nanmedian

→Tensor or tuple of Tensor

nanmedian(input: Tensor, dim: _int | None = ..., keepdim: _bool = ...)

source edit

Median of a tensor, ignoring NaN entries.

Computes the median over the requested axis (or the entire tensor when dim is None), treating NaN entries as missing data rather than propagating them. This is the NaN-safe counterpart of median.

Parameters

inputTensor

Input tensor of a floating or integer dtype.

dimint | None

Reduction axis. None (default) flattens input first.

keepdimbool

If True, retain the reduced axis as a singleton in the output shape. Ignored when dim is None. Defaults to False.

Returns

Tensor or tuple of Tensor

With dim is None, the scalar median of the non-NaN entries. With dim given, (values, indices): the median of every slice along dim and the int64 position of that median in input, both the input shape with dim removed (or kept as size 1 if keepdim=True).

Raises

NotImplementedError

For a bool or complex input.

IndexError

When dim is out of range, or the reduced axis has length 0.

Notes

Every slice keeps its own count $n$ of non-NaN entries. NaN sorts last, so the median is the entry at sorted position $\lfloor (n - 1) / 2 \rfloor$ — the lower median for an even count, matching the reference framework. A slice that is all NaN has no number to take and answers NaN, at the position of its first entry. Integer inputs hold no NaN and take the plain lower median.

The gradient of a dim reduction flows to the entry indices names. The whole-tensor median shares its gradient evenly between every entry equal to it (every NaN, when the answer is NaN), as the reference does.

Examples

>>> import lucid
>>> import math
>>> x = lucid.tensor([1.0, math.nan, 3.0, 5.0])
>>> lucid.nanmedian(x)
tensor(3.)
>>> values, indices = lucid.nanmedian(
...     lucid.tensor([[1.0, math.nan, math.nan], [1.0, 2.0, 3.0]]), 1
... )
>>> values
tensor([1., 2.])
>>> indices
tensor([0, 1], dtype=lucid.int64)

>>> import lucid >>> import math >>> x = lucid.tensor([1.0, math.nan, 3.0, 5.0]) >>> lucid.nanmedian(x) tensor(3.) >>> values, indices = lucid.nanmedian( ... lucid.tensor([[1.0, math.nan, math.nan], [1.0, 2.0, 3.0]]), 1 ... ) >>> values tensor([1., 2.]) >>> indices tensor([0, 1], dtype=lucid.int64)