In [1]:

    

import numpy as np
import matplotlib.pylab as plt
import pymc3 as pm
import seaborn as sns

from warnings import filterwarnings

filterwarnings('ignore')

sns.set_style('whitegrid')

print('Running on PyMC3 v{}'.format(pm.__version__))

%matplotlib inline


    

    




Running on PyMC3 v3.6

In [5]:

    

totals = np.array([10123, 11098, 9879, 10200, 9898, 9987])
amounts = np.array([800, 900, 1100, 1019, 879, 860])


    

In [6]:

    

p_obs = amounts/totals


    

In [7]:

    

np.allclose(totals * p_obs, amounts)


    

    
Out[7]:





True

In [8]:

    

with pm.Model() as model:

    a = pm.HalfNormal('a', .5) #0.5)
    b = pm.HalfNormal('b', .5) #0.1)
    
    p = pm.Beta("p", alpha=a, beta=b)
    
    l = totals * p
    
    s = pm.Poisson("spend", 
                       mu=l, 
                       observed=amounts
                  )


    

In [9]:

    

with model:
    trace = pm.sample(draws=2000)


    

    




Auto-assigning NUTS sampler...
Initializing NUTS using jitter+adapt_diag...
Multiprocess sampling (4 chains in 4 jobs)
NUTS: [p, b, a]
Sampling 4 chains: 100%|██████████| 10000/10000 [00:04<00:00, 2374.22draws/s]
There were 3 divergences after tuning. Increase `target_accept` or reparameterize.

In [10]:

    

pm.plot_posterior(trace, varnames=["p"], ref_val=p_obs.mean())


    

    
Out[10]:





<matplotlib.axes._subplots.AxesSubplot at 0x11f565be0>






    

In [18]:

    

with model:
    ppc = pm.sample_posterior_predictive(trace)


    

    




100%|██████████| 8000/8000 [00:03<00:00, 2390.79it/s]

In [19]:

    

ppc["spend"].shape


    

    
Out[19]:





(8000, 6)

In [11]:

    

with pm.Model() as model2:

    a = pm.HalfNormal('a', .5) #0.5)
    b = pm.HalfNormal('b', .5) #0.1)
    
    p = pm.Beta("p", alpha=a, beta=b, shape=len(p_obs))
    
    l = totals * p
    
    s = pm.Poisson("spend", 
                       mu=l, 
                       observed=amounts
                  )


    

In [12]:

    

with model2:
    trace2 = pm.sample(draws=2000)


    

    




Auto-assigning NUTS sampler...
Initializing NUTS using jitter+adapt_diag...
Multiprocess sampling (4 chains in 4 jobs)
NUTS: [p, b, a]
Sampling 4 chains: 100%|██████████| 10000/10000 [00:04<00:00, 2326.25draws/s]

In [14]:

    

pm.plot_posterior(trace2, varnames=["p"], ref_val=p_obs);


    

In [15]:

    

with model2:
    ppc2 = pm.sample_posterior_predictive(trace2)


    

    




100%|██████████| 8000/8000 [00:02<00:00, 2824.54it/s]

In [17]:

    

ppc2["spend"].shape


    

    
Out[17]:





(8000, 6)

In [ ]: