[discussion] Parameters not rendering when feeding into pyro.deterministic

[KevinTran-TRI]

Issue Description

Parts of my programs are deterministic, and these parts still have latent parameters I would like to make inferences of. But when I try to render my program, my registered parameters are not showing up. It looks like this only happens when these parameters are feeding strictly deterministic variables.

I suspect they're not rendering because parameters that feed only into deterministic "variables" will not actually be part of optimization. Is this true?

Environment

• OS: macOS Sonoma 14.5
• Python version: 3.11.4
• PyTorch version: 2.0.0
• Pyro version: 1.9.0+f02dfb9

Code Snippet

import pyro
import pyro.distributions as dist
import torch


data = torch.tensor([1.0, 2.0, 3.0])

def deterministic_model(data):
    value = pyro.param("param", torch.tensor(0.))
    with pyro.plate("plate", len(data)):
        pyro.deterministic("deterministic", data + value)

pyro.render_model(deterministic_model, model_args=(data,), render_params=True)

def probabilistic_model(data):
    value = pyro.param("param", torch.tensor(0.))
    with pyro.plate("plate", len(data)):
        pyro.sample("probabilistic", dist.Normal(value, 1), obs=data)

pyro.render_model(probabilistic_model, model_args=(data,), render_params=True)

[BenZickel]

Hi @KevinTran-TRI, the behavior you are seeing is a bug only in rendering the model's graph and will be fixed by #3378.

I'm not sure what you are trying to do but the computation is done correctly.

Please note however that deterministic nodes are used mainly for recording values which are completely determined by their parents and they do not affect the model probability density and cannot be observed.