log1p_op

→TensorImplPtr

int log1p_op(const int & a)

source edit

Python wrappers

pylucid.ops.log1pfunction

Natural log of 1 + x: $y = \ln(1 + x)$ .

Composite over add_op + log_op. Inherits the underlying log_op precision — does not provide the high-accuracy small- $x$ behaviour of a dedicated log1p kernel. Gradient flows through LogBackward and AddBackward.

Math

y = \ln(1 + x), \qquad \frac{\partial y}{\partial x} = \frac{1}{1 + x}

See Also

log_op, log10_op.

Parameters

aTensorImplPtr

Input tensor. Domain:

x > -1

Returns

TensorImplPtr

Tensor of the same shape and dtype as a.

Notes

For $|x| \ll 1$ accuracy degrades relative to libm's log1p — catastrophic cancellation happens before the log is applied.