nan_to_num

→Tensor

nan_to_num(input: Tensor, nan: float | None = ..., posinf: float | None = ..., neginf: float | None = ...)

source

Replace NaN, positive and negative infinity entries with finite values.

Returns a tensor identical to input except that NaN, +inf and -inf entries are substituted by nan, posinf and neginf respectively. When a replacement argument is None the engine picks a sensible default (typically zero for NaN and the dtype-specific finite extrema for the infinities).

Parameters

inputTensor

Input tensor.

nanfloat or None= ...

Value used to replace NaN entries.

posinffloat or None= ...

Value used to replace +inf entries.

neginffloat or None= ...

Value used to replace -inf entries.

Returns

Tensor

Sanitized tensor with the same shape and dtype as input.

Notes

Useful as a numerical safety net after operations that can produce non-finite intermediate values (logs, divisions, large exponentials). Gradient flows through the substitution unchanged.

Examples

>>> import lucid
>>> x = lucid.tensor([1.0, float('nan'), float('inf'), float('-inf')])
>>> lucid.nan_to_num(x, nan=0.0, posinf=1e10, neginf=-1e10)
Tensor([1.0, 0.0, 1e10, -1e10])