Symmetric vs Asymmetric Quant Error

Problem

Compare MSE of symmetric and asymmetric int quantization. Symmetric uses scale = max(|x|) / (2^(bits-1)-1) with codes clamped to +/-qmax; asymmetric uses step = (max-min)/(2^bits-1) and the zero point implied by round((x-min)/step). Return [sym_mse, asym_mse].

The signature is symmetric_vs_asymmetric_mse(values, bits=8).

Hint

Asymmetric quantization uses the full code range but pays for a zero point.

Test cases

case 1

in: [[-1,1],8]

expected: [0,0]

case 2

in: [[-1,0,0.5],8]

expected: [0.000005166677000020648,0]

case 3

in: [[-3,4,1],4]

expected: [0.013605442176870777,0.013333333333333357]

case 4

in: [[0,0],8]

expected: [0,0]

Starter codePython

def symmetric_vs_asymmetric_mse(values, bits=8):
    # Your code here
    pass

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