Basic getting started question - how to condition a model

1

I’m trying to get started learning Pyro. The documentation looks very good but it’s also quite confusing near the beginning.

So I’m working with one of the first models introduced, the weather one. Here’s my version of it:

def weather():
    sky = pyro.sample("sky",
        pyro.distributions.Bernoulli(0.3)
    )
    sky = {
        0.0: "cloudy",
        1.0: "sunny"
    }[sky.item()]
    temp_mean = {"cloudy": 55.0, "sunny": 75.0}[sky]
    temp_scale = {"cloudy": 10.0, "sunny": 15.0}[sky]
    temp = pyro.sample("temp",
        pyro.distributions.Normal(temp_mean,temp_scale)
    )
    return sky, temp

I can create a conditioned version like this

conditioned_weather = pyro.condition(
    weather, 
    data={"temp": torch.tensor(80)}
)

but I don’t now know how to actually evaluate this. Calling conditioned_weather returns either ('sunny', tensor(80)) or ('cloudy', tensor(80)), but not with the high frequency of sunny outcomes that I might expect.

I understand that the reason for this is that I need to create a “guide” (variational distirbution) and optimise that, but I’m having trouble seeing how to do that because the examples all seem to be doing something more complicated, with multiple observations instead of a single conditioning step. I’m also not sure what the guide should be in this case, because I don’t really want it to be different from the model (or at least I don’t think I do) - I really just want to find the conditional distribution.

Would someone be able to work me through the basic steps of how to do this simple task with Pyro?

2

Hi @n_virgo, I don’t quite understand your question, but I agree that the introductory tutorial is confusing and have written a completely new one that is now available on the examples web page. You might try looking through that to see if it helps clear things up.

3

Thank you @eb8680_2, I’ve read through most of the updated tutorial now and it’s a huge improvement, enormously clearer than the previous version.

1 Like