cdist

→Tensor

cdist(x1: Tensor, x2: Tensor, p: float = 2.0)

source edit

Pairwise $L_p$ distance matrix between rows of x1 and x2.

Specialised paths exist for p in $\{1, 2, \infty, 0\}$ ; other p values fall back to the general $(\sum_k |a_k - b_k|^p)^{1/p}$ formula. For p=2 the numerically stable expansion $\|a-b\|^2 = \|a\|^2 + \|b\|^2 - 2 a \cdot b^\top$ is used so large coordinates don't lose precision.

Parameters

x1Tensor

Shape (..., P, M).

x2Tensor

Shape (..., R, M). Leading batch dimensions are broadcast-compatible with x1.

pfloat= 2.0

L_p

exponent. Default 2.0 (Euclidean). Use float("inf") for Chebyshev, 0.0 for the Hamming-style non-zero count.

Returns

Tensor

Shape (..., P, R). Entry $(\ldots, i, j)$ = $\|x_1[\ldots, i, :] - x_2[\ldots, j, :]\|_p$ .

Raises

DtypeMismatch

If x1 or x2 is not a real floating-point tensor, as the reference refuses integer, bool and complex inputs — a TypeError and a RuntimeError.