Per-Sample Grad Norm

Problem

Compute the Euclidean norm of a single per-sample gradient.

Return sqrt(sum(g ** 2)); an empty gradient returns 0.0. This value feeds per-sample clipping in differentially private training.

Hint

The norm is the Euclidean length of the gradient vector.

case 1

in: [[3,4]]

expected: 5

case 2

in: [[0,0]]

expected: 0

case 3

in: [[1,-1,1]]

expected: 1.7320508075688772

case 4

in: [[]]

expected: 0

Starter codePython

def per_sample_grad_norm(grad):
    # Your code here
    pass

4 test cases · run your code against them in the editor