Consider that I have a neural network-based guide to learn how to do inference in a pyro generative program. The thing is, some layers of my guide are not supposed to be optimized. I am currently approaching this by not registering in the Pyro this
nn.module as a
pyro.module. Imagine that my
self.g2 neural modules; if at the beginning of my
forward() function I only register g1, i.e.
pyro.module("g1", self.g1), it means my optimizer won’t update the parameters of
self.g2 right? Thank you in advance!