NotImplementedError: Cannot transform _IntegerGreaterThan constraints

1

Edit:
I’m using pyro-ppl-0.2.1+8a9f0e68 and pytorch-0.4.1

Hi,

I’m trying to recreate a question from “Probabilistic Programming and Bayesian Methods for Hackers” (10/10 would recommend). The problem’s link is given below

Basically, the task is to model rate of messages in a day, where on some day the rate changes for some reason. I used two dist.Exponentials to sample lambdas for two dist.Poissons and a dist.Uniform to sample a day after which model switchs from one lambda to another. data is a global variable having observed messages for 70 odd days. My code looks like this:

class NMessages(TorchDistribution):
    def __init__(self):
        self.l1 = dist.Exponential(1/data.mean())
        self.l2 = dist.Exponential(1/data.mean())
        self.tau = dist.Uniform(0, len(data)+1)
        
    def sample(self, sample_shape=torch.Size([]), obs=None):
        tau = self.tau.sample()
        samples = []
        obs = [None for _ in range(len(data))] if obs is None else obs
        l1 = pyro.sample('l1', self.l1)
        l2 = pyro.sample('l2', self.l2)
        
        for t in pyro.irange("data_loop", len(data)):
            if t < tau:
                samples.append(pyro.sample(f'msg_{t+1}', dist.Poisson(l1), obs=obs[t]))
            else:
                samples.append(pyro.sample(f'msg_{t+1}', dist.Poisson(l2), obs=obs[t]))
        return torch.tensor(samples)

When I try to run MCMC inference using

n = NMessages()

nuts_kernel = NUTS(n, adapt_step_size=True)
mcmc_run = MCMC(nuts_kernel, num_samples=500, warmup_steps=300).run(torch.tensor(data))

I get the error stated in the title. I’m attaching the stack trace as well…

NotImplementedError                       Traceback (most recent call last)
<ipython-input-7-cac768c3181f> in <module>()
      4 
      5 nuts_kernel = NUTS(n, adapt_step_size=True)
----> 6 mcmc_run = MCMC(nuts_kernel, num_samples=500, warmup_steps=300).run(torch.tensor(d))
      7 
      8 def infer(parameter):

~/anaconda3/envs/fastai/lib/python3.6/site-packages/pyro/infer/abstract_infer.py in run(self, *args, **kwargs)
     82         self._init()
     83         with poutine.block():
---> 84             for tr, logit in self._traces(*args, **kwargs):
     85                 self.exec_traces.append(tr)
     86                 self.log_weights.append(logit)

~/anaconda3/envs/fastai/lib/python3.6/site-packages/pyro/infer/mcmc/mcmc.py in _traces(self, *args, **kwargs)
    255 
    256     def _traces(self, *args, **kwargs):
--> 257         for sample in self.sampler._traces(*args, **kwargs):
    258             yield sample

~/anaconda3/envs/fastai/lib/python3.6/site-packages/pyro/infer/mcmc/mcmc.py in _traces(self, *args, **kwargs)
    198             if not is_multiprocessing else None
    199         self.logger = initialize_logger(self.logger, logger_id, progress_bar, log_queue)
--> 200         self.kernel.setup(self.warmup_steps, *args, **kwargs)
    201         trace = self.kernel.initial_trace()
    202         with optional(progress_bar, not is_multiprocessing):

~/anaconda3/envs/fastai/lib/python3.6/site-packages/pyro/infer/mcmc/hmc.py in setup(self, warmup_steps, *args, **kwargs)
    368             site_value = node["value"]
    369             if node["fn"].support is not constraints.real and self._automatic_transform_enabled:
--> 370                 self.transforms[name] = biject_to(node["fn"].support).inv
    371                 site_value = self.transforms[name](node["value"])
    372             self._r_shapes[name] = site_value.shape

~/anaconda3/envs/fastai/lib/python3.6/site-packages/torch/distributions/constraint_registry.py in __call__(self, constraint)
    140         except KeyError:
    141             raise NotImplementedError(
--> 142                 'Cannot transform {} constraints'.format(type(constraint).__name__))
    143         return factory(constraint)
    144 

NotImplementedError: Cannot transform _IntegerGreaterThan constraints

I’m not sure what I’m missing, or doing wrong.

2

The issue is that some of your Poisson variables are not observed, but HMC can only handle discrete variables if they are observed. There are a couple ways to work around this, one is to wrap in a poutine.block:

from pyro.util import optional
from pyro import poutine
...
lt = l1 if t < tau else l2
with optional(poutine.block(), obs[t] is None):
    x = pyro.sample(f'msg_{t+1}', dist.Poisson(lt), obs=obs[t]))
samples.append(x)
1 Like