All problems
Decoupled vs Coupled Decay Delta
Problem
Measure the per-step difference between decoupled and coupled weight decay under an adaptive optimizer.
Coupled decay is scaled by the adaptive denominator, giving lr * wd * w / denom, while decoupled decay gives lr * wd * w. Return decoupled - coupled, where denom is sqrt(v_hat) + eps.
Hint
The adaptive denominator changes how much shrinkage a coupled penalty actually applies.
Test cases
case 1
in: [2,0.1,0.01,2]
expected: 0.001
case 2
in: [1,0.5,0.1,1]
expected: 0
case 3
in: [-3,0.2,0.05,4]
expected: -0.022500000000000006
case 4
in: [0,0.1,0.1,1]
expected: 0
Starter codePython
def decoupled_vs_coupled_delta(w, lr, wd, denom):
# Your code here
pass4 test cases · run your code against them in the editor