Posterior Predictive Distribution using MCMC

Tutorials
1

Dear fellow pyromants,

i have a question regarding how to obtain the posterior predictive distribution using MCMC techniques.
Here is the little example code i am trying to run:

# generate some data
np.random.seed(1)
xs = to.tensor(np.linspace(-2,2,23)).float()
f = lambda z : z*.5 + 1
ys = f(xs) + to.tensor(np.random.normal(0, .5, size=len(xs))).float()

def y_model(x, y_obs=None):
    ''' very simple linreg model'''
    # priors
    w = pyro.sample('w', pyro.distributions.Normal(loc=0, scale=10))
    b = pyro.sample('b', pyro.distributions.Normal(loc=0, scale=10))
    
    f_x = w*x + b
    
    y = pyro.sample('y', pyro.distributions.Normal(loc=f_x, scale=1), obs=y_obs)
    
    return f_x

I can now sample models from the prior by just running
samples_from_prior = y_model(xs, None)

For inference, as described in the title, i want to use MCMC techniques:

from pyro.infer.mcmc import MCMC, NUTS
posterior_trace = MCMC(kernel=NUTS(y_model), num_samples=1337, warmup_steps=100)
posterior_trace.run(xs, ys)

The posterior_trace is now an pyro.infer.mcmc.mcmc object, on which i can call posterior_trace.marginal(sites=..).empirical[..] to get the samples for the respective random variables “w” and “b” (and also ‘y’). And this is about how far the tutorial goes.

But now i would like to sample models y_model given the parameter configurations in the posterior_trace. I am not sure how this is ment to be done.

In case of SVI and a given guide i usually did the following:
model_predictions = []
for _ in range(100): # sample 100 models (i.e. posteriors given y_obs)

## trace of the guide (samples all variables given x from the guide)
guide_trace = pyro.poutine.trace(guide).get_trace(xs, None)

# runs the model using samples from the guide for the latent variables
replay_result = pyro.poutine.replay(y_model, guide_trace)(xs, None)

# append to list
model_predictions.append(replay_result.detach().numpy())

which would give me 100 sampled models from the (guide-approximated) posterior predictive distribution.
How can i achieve the same using the results obtained from MCMC ?

Thanks in advance !

PS:

2

You can use pyro.infer.abstract_infer.TracePredictive to get a representation of the posterior predictive distribution from an MCMC run:

posterior = MCMC(NUTS(model), ...).run(xs, ys)
posterior_predictive = TracePredictive(model, posterior, num_samples=100)
posterior_sample = posterior_predictive(xs)  # a Trace object

See examples of the use of TracePredictive in the baseball example.

1 Like
3

First of all, thank you for you help ! It is greatly appreciated !

I tried to run your proposed code but it did not work, i received the following error:

from pyro.infer import TracePredictive
from pyro.infer.mcmc import MCMC, NUTS

posterior = MCMC(kernel=NUTS(y_model), num_samples=1337, warmup_steps=500).run(xs, ys)

posterior_predictive = TracePredictive(y_model, posterior, num_samples=100)
posterior_sample = posterior_predictive(xs)  # a Trace object

Sample: 100%|█| 1837/1837 [00:33<00:00, 55.51it/s, step size=1.12e+00, acc. rate=0.892]
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-40-ce761d5a05f9> in <module>
      5 
      6 posterior_predictive = TracePredictive(y_model, posterior, num_samples=100)
----> 7 posterior_sample = posterior_predictive(xs)  # a Trace object

~\Anaconda3\lib\site-packages\pyro\infer\abstract_infer.py in __call__(self, *args, **kwargs)
    203         # we get the index from ``idxs_by_chain`` instead of sampling from
    204         # the marginal directly.
--> 205         random_idx = self._categorical.sample().item()
    206         chain_idx, sample_idx = random_idx % self.num_chains, random_idx // self.num_chains
    207         sample_idx = self._idx_by_chain[chain_idx][sample_idx]

AttributeError: 'NoneType' object has no attribute 'sample'

I also tried to look into the Baseball example, but it is a little to comprehensive for me to understand yet. I am still struggling with wrapping my head around examples that go over a few lines.

I managed to “sample models” in hacky way through pyro.poutine.condition:

sampled_parameters = posterior.marginal(sites=['w', 'b']).empirical

model = pyro.poutine.condition(y_model, 
                               data=dict((k, v._samples.reshape(-1, 1)) 
                                         for k, v in sampled_parameters.items()))

But for this approach i have to provide the sites of all the parameters manually.

4

It seems like the posterior samples are not yet populated and hence you see this error. Try the following (.run would generate and store forward samples by resampling sample sites from the posterior).

posterior_predictive = TracePredictive(y_model, posterior, num_samples=100).run(xs, None)
posterior_sample = posterior_predictive(xs)
# or 
posterior_samples = posterior_predictive.marginal().empirical['_RETURN']

Let me know if that does not work.

1 Like
5

Thank you very much !

from pyro.infer import TracePredictive
from pyro.infer.mcmc import MCMC, NUTS
# 
posterior = MCMC(kernel=NUTS(y_model), num_samples=1337, warmup_steps=500).run(xs, ys)

posterior_predictive = TracePredictive(y_model, posterior, num_samples=100).run(xs, None)

sampled_model_outputs = posterior_predictive.marginal().empirical['_RETURN']._samples

Works like a charm !

But i am not sure how to properly make use of the pyro.poutine.trace_struct.Trace object returned by
posterior_sample = posterior_predictive(xs)

could you elaborate on this ?

6

The trace data structure (which depending on the version of Pyro you are using would either be a networkx.DiGraph instance or simply a Python class that wraps over dicts of nodes and edges in your graphical model), can be used to access the value of any sample site (i.e. sample from the posterior) using trace.nodes[site_name]. Alternatively you could also access the marginal distribution of any site directly, as is done by sampled_model_outputs for the _RETURN site:

posterior_predictive.marginals([site_name]).empirical[site_name]

So depending on what you need you can get away with just the .marginal method or access the execution traces directly and post-process them for your analysis.

7

Hi,

I’m also trying to get posterior predictions out of the mcmc. But when I run the first line of the code you provided, it gives the following error:
AttributeError: ‘NoneType’ object has no attribute ‘exec_traces’

8

Pyro now has two apis for MCMC, and the two doesn’t work the same way… There is pyro.infer.mcmc.api and then there is pyro.infer.mcmc. I got the same error when using the .api interface, but not the other. This is indeed a bit confusing.

9

We will be removing the earlier interface, which is currently only there to provide backwards compatibility. Note that with the api.MCMC you, you can directly use the simpler predictive utility function from pyro.infer.mcmc.util. TracePredictive is also deprecated and will be removed in the next release. Sorry about the confusion - once we remove the old code, things should be much cleaner and easier to discover.

1 Like