Variable total_size in Binomial gives ValueError 'The value argument must be within the support'

1

I have a model which randomly sizes a Binomial and then samples from it. However, it appears to give an error when the total_count ends up less than the observed value. Shouldn’t this just assign this branch a zero probability and continue sampling? The stack trace is below. Thanks,

Ravi

/usr/local/lib/python2.7/dist-packages/pyro/infer/svi.pyc in step(self, *args, **kwargs)
73 # get loss and compute gradients
74 with poutine.trace(param_only=True) as param_capture:
—> 75 loss = self.loss_and_grads(self.model, self.guide, *args, **kwargs)
76
77 params = set(site[“value”].unconstrained()

/usr/local/lib/python2.7/dist-packages/pyro/infer/trace_elbo.pyc in loss_and_grads(self, model, guide, *args, **kwargs)
105 elbo = 0.0
106 # grab a trace from the generator
–> 107 for model_trace, guide_trace in self._get_traces(model, guide, *args, **kwargs):
108 elbo_particle = 0
109 surrogate_elbo_particle = 0

/usr/local/lib/python2.7/dist-packages/pyro/infer/trace_elbo.pyc in _get_traces(self, model, guide, *args, **kwargs)
66 model_trace = prune_subsample_sites(model_trace)
67
—> 68 model_trace.compute_log_prob()
69 guide_trace.compute_score_parts()
70 if is_validation_enabled():

/usr/local/lib/python2.7/dist-packages/pyro/poutine/trace_struct.pyc in compute_log_prob(self, site_filter)
250 except KeyError:
251 args, kwargs = site[“args”], site[“kwargs”]
–> 252 site_log_p = site[“fn”].log_prob(site[“value”], *args, **kwargs)
253 site_log_p = scale_tensor(site_log_p, site[“scale”])
254 site[“log_prob”] = site_log_p

/usr/local/lib/python2.7/dist-packages/pyro/distributions/binomial.pyc in log_prob(self, value)
101 def log_prob(self, value):
102 if self._validate_args:
–> 103 self._validate_sample(value)
104 log_factorial_n = torch.lgamma(self.total_count + 1)
105 log_factorial_k = torch.lgamma(value + 1)

/usr/local/lib/python2.7/dist-packages/torch/distributions/distribution.pyc in _validate_sample(self, value)
219
220 if not self.support.check(value).all():
–> 221 raise ValueError(‘The value argument must be within the support’)
222
223 def repr(self):

ValueError: The value argument must be within the support

2

This is expected, and even desirable because this is typically indicative of errors in your code. If you need to score samples with negligible probability, you should include them in your support (via the probs arg with a value of 0.0 or a small epsilon). That will behave as what you are expecting. Note that for numerical stability reasons, you might not exactly get -inf when you try to call .log_prob on such a sample, but it will be close within the precision type used.

Another thing you could do is disable distribution args/value checking via pyro.distributions.enable_validation(False), which will turn off all validation checks. That will probably give you the result that you are expecting, except that it may lead to unexpected behavior somewhere else (due to NaNs for instance).