MCMC().run() does not return a TracePosterior object

[djinnome]

Hi @robertness

I have been going through the intro_to_generative_ML_with_Pyro.ipynb, and when I got to this code block:

from pyro.infer import NUTS, MCMC, EmpiricalMarginal

guess_prior = torch.tensor(13.)

nuts_kernel = NUTS(conditioned_scale, adapt_step_size=True)
posterior = MCMC(nuts_kernel, 
                 num_samples=1000, 
                 warmup_steps=300).run(guess_prior)
marginal = EmpiricalMarginal(posterior, "weight")

plt.hist([marginal().item() for _ in range(1000)],)
plt.title("P(weight | measurement = 14)")
plt.xlabel("Weight")
plt.ylabel("#")
print(posterior)

I got the following error:

---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
<ipython-input-16-4bbde7897d55> in <module>
----> 1 marginal = EmpiricalMarginal(posterior, "weight")
      2 
      3 plt.hist([marginal().item() for _ in range(1000)],)
      4 plt.title("P(weight | measurement = 14)")
      5 plt.xlabel("Weight")

~/.pyenv/versions/anaconda3-2019.10/lib/python3.7/site-packages/pyro/infer/abstract_infer.py in __init__(self, trace_posterior, sites, validate_args)
     30     def __init__(self, trace_posterior, sites=None, validate_args=None):
     31         assert isinstance(trace_posterior, TracePosterior), \
---> 32             "trace_dist must be trace posterior distribution object"
     33         if sites is None:
     34             sites = "_RETURN"

AssertionError: trace_dist must be trace posterior distribution object

I believe part of the problem is this line:

posterior = MCMC(nuts_kernel, 
                 num_samples=1000, 
                 warmup_steps=300).run(guess_prior)

because MCMC().run() does not appear to return anything, whereas the EmpiricalMarginal() is expecting an instance of TracePosterior.

Please advise.

Sincerely,

Jeremy

[djinnome]

Hi @robertness

OK, I figured out the problem. In Pyro version 0.3.3, the MCMC object is a subclass of TracePosterior

but since 0.3.4, it isn't.

I downgraded to 0.3.3 and now I can compute the EmpiricalMarginal, but when I try to execute:

with pyro.plate("x_axis", 5):
    d = dist.MultivariateNormal(torch.zeros(3), torch.eye(3, 3))
    x = pyro.sample("x", d)
x.shape

The Jupyter kernel restarts.

Not sure how to get past this one. I am running on a Macbook Pro 2015, so maybe I'll try on linux?

Sincerely,

Jeremy

[vvrahul11]

Hi @robertness

OK, I figured out the problem. In Pyro version 0.3.3, the MCMC object is a subclass of TracePosterior

but since 0.3.4, it isn't.

I downgraded to 0.3.3 and now I can compute the EmpiricalMarginal, but when I try to execute:

with pyro.plate("x_axis", 5):
    d = dist.MultivariateNormal(torch.zeros(3), torch.eye(3, 3))
    x = pyro.sample("x", d)
x.shape

The Jupyter kernel restarts.

Not sure how to get past this one. I am running on a Macbook Pro 2015, so maybe I'll try on linux?

Sincerely,

Jeremy

@djinnome You can use get_samples() instead. [pyro-api-0.1.2 pyro-ppl-1.3.1]

from pyro.infer import NUTS, MCMC
guess_prior = torch.tensor(13.)
nuts_kernel = NUTS(conditioned_scale, adapt_step_size=True)
mcmc = MCMC(nuts_kernel, num_samples=1000, warmup_steps=300)
mcmc.run(guess_prior)
posterior_a = mcmc.get_samples(50)

plt.hist(posterior_a['weight'])
plt.title("P(weight | measurement = 14)")
plt.xlabel("Weight")
plt.ylabel("#")
print(posterior)
posterior_a['weight']

`