uniform binning under normal_flow function in model.py

[MichalisLazarou]

Hey,

Many thanks on your nicely written code. I am trying to understand better normalizing flows but I am struggling to understand in the normal_flow function the use of uniform_binning_correction function. As I can see it adds a small offset of 1/256 to every value of x, why is this the case? I assume this is because the pixels are discretized, in case the function is continuous how will normalizing flows work there?

[MaveriQ]

@MichalisLazarou Have you found an intuition or explanation for this function? Thanks

[GitYCC]

You can reference UNIFORM DEQUANTIZATION in this paper.