value_error_test.py

from pymc import *

def test(a0, b0):
	# use non-informative priors
	x = rbinomial(10, .5, 10)
	a = Uniform('a', lower=-1000, upper=1000, value=a0)
	b = Normal('b', mu=0, tau=.01, value=b0)
	
	@deterministic
	def theta(a=a, b=b):
		return a + b * x
	
	p = InvLogit('p', theta)
	b = Binomial('b', n=10, p=p, value=x, observed=True) #, verbose=True)

test(.5, 0) # always works
test(.5, None) # sometimes works
test(None, .5) # sometimes works
test(None, None) # almost never works

Not too surprising, right? Anything larger (smaller) than 5 (-5) on the nominal scale will be at the boundaries of p. So, positive values of x will result in a ZeroProbabilityError for p=0 (and similarly for p=0 when x<n).