openai/glow

why no logdet computed for the actnorm in conv2d

weixsong opened this issue · 1 comments

glow/tfops.py

Line 256 in eaff217

x = actnorm("actnorm", x)

In this line of code, after conv2d, another actnorm is used.
I'm not quite clear why logdet is not computed for this actnorm, this actnorm also changed the distribution volume.

The conv2d operations are not applied to the random variable directly. They are used by the affine coupling layer to parameterize the affine transformation on the random variable.