linalg_vector_norm

Computes the vector \(p\)-norm along one or more axes:

\[\lVert x \rVert_p = \left( \sum_i \lvert x_i \rvert^{p} \right)^{1/p}\]

For ord = +inf / -inf, the sum is replaced by max / min over the reduced axes.

ATen source: aten.linalg_vector_norm

Inputs

Name	Description
input	Input tensor

Attributes

Name	Type	Description
ord	int / float	Order of the norm. +inf / -inf use max / min instead of the sum
axes	list[int]	Dims to reduce. None reduces over all dims
keep_dim	bool	If True, reduced dims remain with size 1; if False, they are removed
version	int	Composite op version

Output

Name	Description
output	Input shape with reduced dims either set to 1 (keep_dim=True) or removed (keep_dim=False)

PyTorch example

import torch

input = torch.randn(8, 128, 256, 1024)
# Note: the PyTorch arg is `dim`; Core AI's IR attribute is `axes`
output = torch.linalg.vector_norm(input, ord=-1.5, dim=[1, 3], keepdim=False)
# Output shape: (8, 256)

Reference

torch.linalg.vector_norm