nan_to_num_op

→TensorImplPtr

int nan_to_num_op(const int & a, double nan_val, double posinf_val, double neginf_val)

source edit

Python wrappers

pylucid.ops.nan_to_numfunction

Replace NaN and $\pm\infty$ values with user-specified finite substitutes, preserving dtype and shape.

Not differentiable: returns a detached tensor with no backward node. Use only in inference / data-cleaning paths.

Math

\text{nan\_val} & x_i \text{ is NaN} \\ \text{posinf\_val} & x_i = +\infty \\ \text{neginf\_val} & x_i = -\infty \\ x_i & \text{otherwise} \end{cases}$$ **Shape** Output shape equals input shape (elementwise). **See Also** `isnan_op`, `isinf_op`, `isfinite_op`.

Parameters

aTensorImplPtr

Input tensor of any floating-point dtype. Integer inputs are passed through unchanged (no non-finite values to replace).

nan_valdouble= None

Replacement for NaN positions. Default 0.0.

posinf_valdouble= None

Replacement for

+\infty

positions. Default is the maximum finite IEEE-754 float32 value (~3.4028e+38).

neginf_valdouble= None

Replacement for

-\infty

positions. Default is the minimum finite IEEE-754 float32 value (~-3.4028e+38).

Returns

TensorImplPtr

New tensor with the same shape, dtype, and device as a; non-finite slots are replaced according to the three scalars.

Notes

The default posinf_val / neginf_val clamp to float32 range even when a is float64. Pass std::numeric_limits<double>::max explicitly to preserve double-precision extremes.

nan_to_num_op

→TensorImplPtr

int nan_to_num_op(const int & a, double nan_val, double posinf_val, double neginf_val)

source edit

Python wrappers

pylucid.ops.nan_to_numfunction

Replace NaN and $\pm\infty$ values with user-specified finite substitutes, preserving dtype and shape.

Not differentiable: returns a detached tensor with no backward node. Use only in inference / data-cleaning paths.

Math

\text{nan\_val} & x_i \text{ is NaN} \\ \text{posinf\_val} & x_i = +\infty \\ \text{neginf\_val} & x_i = -\infty \\ x_i & \text{otherwise} \end{cases}$$ **Shape** Output shape equals input shape (elementwise). **See Also** `isnan_op`, `isinf_op`, `isfinite_op`.

Parameters

aTensorImplPtr

Input tensor of any floating-point dtype. Integer inputs are passed through unchanged (no non-finite values to replace).

nan_valdouble= None

Replacement for NaN positions. Default 0.0.

posinf_valdouble= None

Replacement for

+\infty

positions. Default is the maximum finite IEEE-754 float32 value (~3.4028e+38).

neginf_valdouble= None

Replacement for

-\infty

positions. Default is the minimum finite IEEE-754 float32 value (~-3.4028e+38).

Returns

TensorImplPtr

New tensor with the same shape, dtype, and device as a; non-finite slots are replaced according to the three scalars.

Notes

The default posinf_val / neginf_val clamp to float32 range even when a is float64. Pass std::numeric_limits<double>::max explicitly to preserve double-precision extremes.