LAMB Trust Ratio Clamp

Problem

Compute the LAMB trust ratio and clamp it to a maximum.

Return min(max_ratio, ||w|| / (||u|| + eps)), where w is the weight vector and u is the raw update vector. Clamping stops a single layer from taking an oversized coordinated step.

Hint

Only the ratio above max_ratio is clipped; small ratios pass through.

Test cases

case 1

in: [[3,4],[0,1],1e-8,5]

expected: 4.999999950000001

case 2

in: [[1,1],[2,2],0,10]

expected: 0.5

case 3

in: [[1,1],[0.01,0.01],1e-8,2]

expected: 2

case 4

in: [[0,0],[1,1],1e-8,4]

expected: 0

Starter codePython

def lamb_trust_ratio_clamp(w, u, eps, max_ratio):
    # Your code here
    pass

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