isfinite_op

→TensorImplPtr

int isfinite_op(const int & a)

Element-wise test for finite values — returns a boolean tensor that is True at every position whose input is neither NaN nor $\pm\infty$ .

Not differentiable: no autograd node is wired and the result is always detached.

Math

$y_i = \neg \operatorname{isnan}(x_i) \wedge \neg \operatorname{isinf}(x_i)$

Shape

Output shape equals input shape (elementwise).

See Also

isnan_op, isinf_op.

aTensorImplPtr

Input tensor of any floating-point dtype. Integer inputs always produce True since every integer value is finite.

TensorImplPtr

Boolean tensor with the same shape and device as a; dtype is Dtype::Bool.

Equivalent to ~(isnan(a) | isinf(a)) but dispatched as a single backend kernel for efficiency.