Information_criterion returns empty dict

Misc.
1

Probably a naive question, but I am following the Bayesian Linear Regression tutorial part 2. After running the SVI for n steps with the custom guide, I try to run svi.information_criterion() but when printed I get {}, an empty dictionary. What might be the problem?

2

i believe you need to call the run method before attempting information_criterion()

3

Source code says svi.run() is deprecated and svi.step() should be used. I guess, this is not compatible with svi.information_criterion()?

4

yeah some of that functionality is either deprecated or only partially implemented so tread with care. however i’ll note that this code snippet

import torch

import pyro
import pyro.distributions as dist
import pyro.optim as optim
from pyro.infer import SVI, Trace_ELBO
from pyro.infer.autoguide import AutoNormal


def model():
    mu = pyro.sample("mu", dist.Normal(0, 1))
    log_sigma = pyro.sample("log_sigma", dist.Normal(0, 1))
    with pyro.plate("plate"):
        pyro.sample("kcal", dist.Normal(mu, log_sigma.exp()), obs=torch.ones(7))

guide = AutoNormal(model)

svi = SVI(
    model,
    AutoNormal(model),
    optim.Adam({"lr": 0.05}),
    loss=Trace_ELBO(),
    num_samples=100,
)
for i in range(100):
    svi.step()

posterior = svi.run()
print(posterior.information_criterion())

produces an output like

OrderedDict([('waic', tensor(-14.1256, grad_fn=<SumBackward0>)), ('p_waic', tensor(0.9437, grad_fn=<SumBackward0>))])