Bits per Dimension

Problem

Return the bits-per-dimension score of a generative model from its total negative log-likelihood in nats:

BPD = nll_nats / (num_samples * dims * ln(2))

Hint

Dividing by ln(2) converts nats into bits.

case 1

in: [1000,10,100]

expected: 1.4426950408889636

case 2

in: [693.1471805599454,10,100]

expected: 1.0000000000000002

case 3

in: [0,5,10]

expected: 0

case 4

in: [100,4,25]

expected: 1.4426950408889634

Starter codePython

import math
def bits_per_dimension(nll_nats, num_samples, dims):
    # Your code here
    pass

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