SAM Sharpness Estimate

Problem

Estimate the local sharpness of the loss from a SAM ascent step.

Return (loss_adv - loss_clean) / rho, the loss increase per unit perturbation. If rho is not positive, return 0.0.

Hint

A larger loss increase over the same perturbation means a sharper minimum.

case 1

in: [0.5,0.7,0.1]

expected: 1.9999999999999996

case 2

in: [1,0.9,0.05]

expected: -1.9999999999999996

case 3

in: [2,2,0.2]

expected: 0

case 4

in: [1,1.5,0]

expected: 0

Starter codePython

def sam_sharpness_estimate(loss_clean, loss_adv, rho):
    # Your code here
    pass

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