Render functions of random variables

[Eric-Bradford]

I would like to show variables explicitly using "render_model" that are functions of other variables. I tried using "deterministic" or "Delta distribution" for this purpose, but this seems to not propagate like other distributions, e.g.

def model(data):
    m = pyro.sample("m", dist.Normal(0, 1))
    sd = pyro.sample("sd", dist.Delta(m**2))
    with pyro.plate("N", len(data)):
        pyro.sample("obs", dist.Normal(m, sd), obs=data)
data = torch.ones(10)
pyro.render_model(model, model_args=(data,))

If I define sd as a normal distribution it shows the connection between sd and obs correctly, but with "Delta" the connection is not shown. How can I define sd as a function of m that would show this connection?

[fritzo]

Hi @Eric-Bradford this looks like a bug to me, let me take a look at what might be causing it...

[Eric-Bradford]

Hi @fritzo, that would be great thanks!

[fritzo]

OK, this is not a bug, merely a historical design choice (pyro.infer.inspect originated as an inference tool that was concerned with only latent variables, hence it ignored deterministic dependencies). I've started sketching support for this feature in a branch but it's a little more complex than I expected due to low-level provenance tracking issues. Working on it...

[fritzo]

See also this forum discussion

[r3v1]

Have been any advances on this?

[fritzo]

I believe there has been no progress, but @eb8680 was looking into it around 6 months ago.

[r3v1]

Is there any plan to implement this in the short time? I've tried patching it somehow but I got lost...

[eb8680]

@r3v1 no, sorry, it's unlikely I/we will have the bandwidth to fix this ourselves any time soon. You're welcome to play around with the branch linked above and see if you can get it working yourself.

[r3v1]

Thanks anyway. I tried but it was quite difficult for me to understand what is going on behind the scenes, I would take it more deeply if I had some documentation, by the way, do you have any resources about?