CFG Variance Rescale

Problem

Rescale a guided prediction to match the conditional prediction's standard deviation (CFG rescaling).

Compute the population standard deviations of cond and guided, then multiply guided elementwise by std_cond / std_guided, rounded to 4 decimal places. If the guided deviation is zero, return guided rounded unchanged.

Test cases

case 1

in: [[1,-1],[3,0]]

expected: [2,0]

case 2

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

expected: [1,1]

case 3

in: [[0,0,0],[5,-5,5]]

expected: [0,0,0]

case 4

in: [[1,2,3],[2,4,6]]

expected: [1,2,3]

Starter codePython

def cfg_rescale(cond, guided):
    # Your code here
    pass

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