Multinomial gives error: ValueError: Stochastic multi: Failed to cast initial value to required dtype.

I am following the book Probabilistic Graphical Model with Python (https://www.packtpub.com/big-data-and-business-intelligence/building-probabilistic-graphical-models-python), trying to run a Bayesian Parameter Estimation. The code for the facility functions is as follows:

def make_model(data,num_draws): # 2-val case (e.g. Intelligence) is done by d   
    partial_dirich = pymc.Dirichlet(name='partial_dirich',theta=[1.,1.,1.]) # k-1 probs stored.
    full_dirich = pymc.CompletedDirichlet(name='full_dirich',D=partial_dirich) # last prob stored.
    multi = pymc.Multinomial(name='multi',value=data,n=num_draws,p=full_dirich,observed=True)
    model = Model([partial_dirich,full_dirich,multi])
    return model

def run_mcmc(model, **kwargs):
    mcmc = pymc.MCMC(model)
    mcmc.sample(**kwargs)
    return mcmc

def get_counts(vals): # from 0,1,2,... result to c(0),c(1),... freq-counts.
    freqs = defaultdict(int)
    for val in vals:
        freqs[val] += 1 # alternatively, b[val] = b.get(val,0) + 1, where get(val,0) means
                        #  "return val if it exists in the dict as a key, otherwise 0."
    return np.asarray(freqs.values())

def get_samples(experiments, evidence, num_samples=10):
    # for n experiments, sample and add the freqs obtained.
    ret = []
    for i in xrange(experiments):
        vals = [float(i['Grade']) for i in bn.randomsample(num_samples,evidence)]
        ret.append(get_counts(vals))
    return ret

def plot_traces(traces):
    # colors = ["#348ABD","#A60628","#884732"]
    colors = ['r','b','g']
    plt.plot(np.arange(len(traces)),traces[:,:,0],c=colors[0])
    plt.plot(np.arange(len(traces)),traces[:,:,1],c=colors[1])
    plt.plot(np.arange(len(traces)),traces[:,:,2],c=colors[2])
    plt.title("Traces of Posteriors")
    plt.show()

def estimate_params(evidence, **kwargs):
    experiments = 10
    samples = get_samples(experiments, evidence, num_samples=10)
    model = make_model(samples, experiments)
    mc = run_mcmc(model,**kwargs)
    traces = mc.trace('full_dirich')[:]
    return [np.mean(traces[:,:,0]), np.mean(traces[:,:,1]), np.mean(traces[:,:,2])], traces

I then ran the following code:

cond_assignments = [['0','0'],['0','1'],['1','0'],['1','1']]
# cond_assignments = [[0,0],[0,1],[1,0],[1,1]]
ret = []
for i,j in cond_assignments:
    evidence = dict(Intelligence=i, Difficulty=j)
    means, traces = estimate_params(evidence, iter=500, burn=100)
    ret.append([[i,j]] + means)
pd.DataFrame(ret, columns=['cond_assignment',0,1,2])

and got the error:

 [-----------------100%-----------------] 500 of 500 complete in 0.1 sec<class 'pymc.distributions.Dirichlet'>
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-664-23e0d4362f67> in <module>()
      4 for i,j in cond_assignments:
      5     evidence = dict(Intelligence=i, Difficulty=j)
----> 6     means, traces = estimate_params(evidence, iter=500, burn=100)
      7     ret.append([[i,j]] + means)
      8 pd.DataFrame(ret, columns=['cond_assignment',0,1,2])

<ipython-input-659-c4ed1fa1bdff> in estimate_params(evidence, **kwargs)
     39     experiments = 10
     40     samples = get_samples(experiments, evidence, num_samples=10)
---> 41     model = make_model(samples, experiments)
     42     mc = run_mcmc(model,**kwargs)
     43     traces = mc.trace('full_dirich')[:]

<ipython-input-659-c4ed1fa1bdff> in make_model(data, num_draws)
      3     full_dirich = pymc.CompletedDirichlet(name='full_dirich',D=partial_dirich) # last prob stored.
      4     print type(partial_dirich)
----> 5     multi = pymc.Multinomial(name='multi',value=data,n=num_draws,p=full_dirich,observed=True)
      6     model = Model([partial_dirich,full_dirich,multi])
      7     return model

/Users/jacobsw/anaconda2/lib/python2.7/site-packages/pymc/distributions.pyc in __init__(self, name, n, p, trace, value, rseed, observed, cache_depth, plot, verbose, **kwds)
   3269                                 rseed=rseed, observed=observed,
   3270                                 cache_depth=cache_depth, plot=plot,
-> 3271                                 verbose=verbose, **kwds)
   3272 
   3273 

/Users/jacobsw/anaconda2/lib/python2.7/site-packages/pymc/PyMCObjects.pyc in __init__(self, logp, doc, name, parents, random, trace, value, dtype, rseed, observed, cache_depth, plot, verbose, isdata, check_logp, logp_partial_gradients)
    749                 'Stochastic %s: Failed to cast initial value to required dtype.\n\nOriginal error message:\n' %
    750                 name + inst.message)
--> 751             six.reraise(cls, new_inst, tb)
    752 
    753         # Store the shape of the stochastic value

/Users/jacobsw/anaconda2/lib/python2.7/site-packages/pymc/PyMCObjects.pyc in __init__(self, logp, doc, name, parents, random, trace, value, dtype, rseed, observed, cache_depth, plot, verbose, isdata, check_logp, logp_partial_gradients)
    740                 # Convert Pandas DataFrames and Series to numpy arrays
    741                 value = getattr(value, 'values', value)
--> 742                 self._value = np.array(value, dtype=dtype)
    743                 self._value.flags['W'] = False
    744             else:

ValueError: Stochastic multi: Failed to cast initial value to required dtype.

Original error message:
setting an array element with a sequence.

Curiously, it works for one run occasionally, but doesn't when I ran multiple experiments. Could anybody help with the issue? Appreciate it.

matthew-brett / pymc

Multinomial gives error: ValueError: Stochastic multi: Failed to cast initial value to required dtype. #1