Why are we sampling z from the standard normal distribution instead of the learned p(z)?

[icbcbicc]

Hi rosinality,

Your code is clean and easy to read. Thank you for your effort.

I have one question: During the sampling process, why are we sampling z from the standard normal distribution (with temperature)? Shouldn't we sample from the learned p(z)? Is it because p(z) is dependant on the data so that we cannot sample from it? (In the implementation, if I'm understanding it correctly, p(z) has four components, three of them are dependent on both the data and the model, while the last one is only dependant on the model.)

Thanks.

[rosinality]

Learned priors used in the sampling step. Please refer to this:

glow-pytorch/model.py

Line 309 in 97081ff

if self.split:

[icbcbicc]

Thanks for your reply!