Question about reporting likelihoods in bits per dim

[Mattias421]

Hi all,

Thank you for providing this code, it's very educational :)

I am interested in reporting likelihoods but I don't fully understand how bits per dim are calculated. In the code this is

bpd = -(prior_logp + delta_logp) / np.log(2)
N = np.prod(shape[1:])
bpd = bpd / N
# A hack to convert log-likelihoods to bits/dim
offset = 7. - inverse_scaler(-1.)
bpd = bpd + offset

Would you be able to elaborate on how this hack works and if it applies to other image dimensions?

Best, Matt

[toyot-li]

same question here.

[mareikethies]

same question here as well Edit: I think this hack applies to images which originally have discrete pixel intensities from 0 to 255. See, e.g., equations 3 to 5 in Danihelka et al. (https://arxiv.org/pdf/1705.05263.pdf).

[toyot-li]

@mareikethies Thanks a lot! I got it.

[Mattias421]

@mareikethies Thank you, it looks like this offset doesn't apply to other data types :0