I am using pyro.infer.SVI for amortized inference.

In this case, isn’t guide q_\phi(z|x)?

My guide consumes data x and gives samples of the latents z to the model.

I am using pyro.infer.SVI for amortized inference.

In this case, isn’t guide q_\phi(z|x)?

My guide consumes data x and gives samples of the latents z to the model.

what is this question in regards to? it’s common to elide dependencies in conditional probability statements if it’s clear from the context otherwise you might get `q_phi(z|x,num_layers,alpha,beta,gamma,delta,xi,rho,sigma,tau)`

This question is regards to the documentation, e.g.: SVI Part III: ELBO Gradient Estimators — Pyro Tutorials 1.8.1 documentation

Since x was included with p, not including the x dependency in q, when it is dependent, could be confusing.

the svi tutorial part iii is agnostic to amortization (it could be amortized; it could not be—but it’s not really important in this context). where amortization is assumed the notation is more explicit

1 Like