def construct_net(self, df):
anxiety = make_bernoulli('anxiety', value=df.anxiety)
peer_pressure = make_bernoulli('peer_pressure', value=df.peer_pressure)
smoking = cartesian_child('smoking',
parents=[anxiety, peer_pressure],
value=df.smoking)
yellow_fingers = cartesian_child('yellow_fingers',
parents=[smoking],
value=df.yellow_fingers)
genetics = make_bernoulli('genetics', value=df.genetics)
lung_cancer = cartesian_child('lung_cancer',
parents=[smoking, genetics],
value=df.lung_cancer)
allergy = make_bernoulli('allergy', value=df.allergy)
coughing = cartesian_child('coughing',
parents=[allergy, lung_cancer],
value=df.coughing)
fatigue = cartesian_child('fatigue',
parents=[lung_cancer, coughing],
value=df.fatigue)
attention_disorder = cartesian_child('attention_disorder',
parents=[genetics],
value=df.attention_disorder)
car_accident = cartesian_child('car_accident',
parents=[fatigue, attention_disorder],
value=df.car_accident)
# sample from the prior
model = pymc.Model([
anxiety, peer_pressure, smoking, yellow_fingers, genetics,
lung_cancer, allergy, coughing, fatigue, attention_disorder,
car_accident
])
return model
def test_cartesian_bernoulli_child_of_categorical_parent():
coeffs = {
'CF: p(feeling_sick)': 0.55,
'CF: p(day_of_week)': [0.013, 0.626, 0.039, 0.108, 0.134, 0.019],
'CF: p(staying_home | day_of_week=5 feeling_sick=1)': 0.240,
'CF: p(staying_home | day_of_week=3 feeling_sick=1)': 0.467,
'CF: p(staying_home | day_of_week=4 feeling_sick=1)': 0.603,
'CF: p(staying_home | day_of_week=0 feeling_sick=0)': 0.974,
'CF: p(staying_home | day_of_week=6 feeling_sick=1)': 0.331,
'CF: p(staying_home | day_of_week=6 feeling_sick=0)': 0.009,
'CF: p(staying_home | day_of_week=2 feeling_sick=1)': 0.317,
'CF: p(staying_home | day_of_week=0 feeling_sick=1)': 0.900,
'CF: p(staying_home | day_of_week=2 feeling_sick=0)': 0.651,
'CF: p(staying_home | day_of_week=1 feeling_sick=0)': 0.603,
'CF: p(staying_home | day_of_week=1 feeling_sick=1)': 0.954,
'CF: p(staying_home | day_of_week=3 feeling_sick=0)': 0.856,
'CF: p(staying_home | day_of_week=5 feeling_sick=0)': 0.606,
'CF: p(staying_home | day_of_week=4 feeling_sick=0)': 0.828
}
np.random.seed = 1
# define the model with fixed *coefficients*
day_of_week = make_categorical('day_of_week', levels=7, N=1, fixed=coeffs)
feeling_sick = make_bernoulli('feeling_sick', N=1, fixed=coeffs)
staying_home = cartesian_bernoulli_child('staying_home',
[day_of_week, feeling_sick],
N=1,
fixed=coeffs)
model = pymc.Model([day_of_week, feeling_sick, staying_home])
# sample data from the model
sampler = pymc.MCMC(model)
sampler.sample(iter=500)
day_of_week_val = sampler.trace('day_of_week')[:]
feeling_sick_val = sampler.trace('feeling_sick')[:]
staying_home_val = sampler.trace('staying_home')[:]
# define the model again this time with fixed *data*
day_of_week, c1 = make_categorical('day_of_week',
levels=7,
value=day_of_week_val,
return_coeffs=True)
feeling_sick, c2 = make_bernoulli('feeling_sick',
value=feeling_sick_val,
return_coeffs=True)
staying_home, c3 = cartesian_bernoulli_child('staying_home',
[day_of_week, feeling_sick],
value=staying_home_val,
return_coeffs=True)
model = pymc.Model([day_of_week, feeling_sick, staying_home] + c1 + c2 +
c3)
sampler = pymc.MCMC(model)
sampler.sample(iter=500, burn=300)
cname = 'CF: p(staying_home | day_of_week=1 feeling_sick=0)'
print cname
assert np.isclose(coeffs[cname], sampler.trace(cname)[:].mean(), atol=0.3)
for c in coeffs:
print c, coeffs[c], sampler.trace(c)[:].mean()
def construct_net(self, df):
rain = make_bernoulli('rain', value=df.rain)
sprinkler = make_bernoulli('sprinkler', value=df.sprinkler)
sidewalk = cartesian_child('wet_sidewalk',
parents=[rain, sprinkler],
value=df.wet_sidewalk)
model = pymc.Model([rain, sprinkler, sidewalk])
return model
def test_cartesian_categorical_child_creates_correctly_named_coefficients():
mum = make_bernoulli('mum', N=1)
dad = make_bernoulli('dad', N=1)
child, coeffs = cartesian_categorical_child('child', [mum, dad],
levels=4,
N=1,
return_coeffs=True)
coeff_names = {str(coeff) for coeff in coeffs}
assert coeff_names == {
'CF: p(child | mum=0 dad=0)', 'CF: p(child | mum=1 dad=1)',
'CF: p(child | mum=1 dad=0)', 'CF: p(child | mum=0 dad=1)', 'p(child)'
}
def construct_net(self, df):
# This is the same network as was used to generate the data, but there
# are some differences in how it is set up:
# - we don't fix the values of coefficients - the whole point is for the model to learn them!
# - we provide 'value' parameter - because these variables are observed
# - we don't need the 'N' parameter - because it is assumed that N = len(value)
rain = make_bernoulli('rain', value=df.rain)
sprinkler = make_bernoulli('sprinkler', value=df.sprinkler)
grass_wet = cartesian_child('grass_wet',
parents=[rain, sprinkler],
levels=2,
value=df.grass_wet)
model = pymc.Model([rain, sprinkler, grass_wet])
return model
def test_make_bernoulli_returns_variable_with_beta_parent():
name = 'rosebud'
bernoulli_var = make_bernoulli(name, value=[0, 1, 1, 0])
parent = bernoulli_var.parents['p']
assert bernoulli_var.observed
assert not parent.observed
assert str(parent) == COEFFS_PREFIX + 'p(%s)' % name
assert isinstance(parent, pymc.distributions.Beta)
def test_cartesian_bernoulli_child_creates_correctly_named_coefficients():
mum = make_bernoulli('mum', N=1)
dad = make_bernoulli('dad', N=1)
child, coeffs = cartesian_bernoulli_child('child', [mum, dad],
N=1,
return_coeffs=True)
coeff_names = {str(coeff) for coeff in coeffs}
assert coeff_names == {
'CF: p(child | mum=0 dad=0)', 'CF: p(child | mum=1 dad=1)',
'CF: p(child | mum=1 dad=0)', 'CF: p(child | mum=0 dad=1)', 'p(child)'
}
single_parent = make_bernoulli('single_parent', N=1)
child, coeffs = cartesian_bernoulli_child('child', [single_parent],
N=1,
return_coeffs=True)
coeff_names = {str(coeff) for coeff in coeffs}
assert coeff_names == {
'p(child)', 'CF: p(child | single_parent=0)',
'CF: p(child | single_parent=1)'
}
def get_rsw_data_random_coeffs(size):
"""generates a data from the famous 'rain - sprinkler - wet grass' bayesian network
returns as a dataframe.
"""
rain = make_bernoulli('rain', N=1)
sprinkler = cartesian_child('sprinkler', parents=[rain], levels=2, N=1)
grass_wet = cartesian_child('grass_wet',
parents=[sprinkler, rain],
levels=2,
N=1)
# sample from the prior
model = pymc.Model([rain, sprinkler, grass_wet])
coefficients = sample_coeffs(model)
# define the bayesian net
rain = make_bernoulli('rain', N=1, fixed=coefficients)
sprinkler = cartesian_child('sprinkler',
parents=[rain],
levels=2,
N=1,
fixed=coefficients)
grass_wet = cartesian_child('grass_wet',
parents=[sprinkler, rain],
levels=2,
N=1,
fixed=coefficients)
# sample from the prior
model = pymc.Model([rain, sprinkler, grass_wet])
sampler = pymc.MCMC(model)
sampler.sample(iter=size + 1000, burn=1000)
data = pd.DataFrame({
str(node): sampler.trace(str(node))[:].ravel() + 0
for node in sampler.stochastics if not str(node).startswith('CF: ')
})
return data
def test_cartesian_bernoulli_child():
# define the model with no data just to sample all the coefficients from
# their priors
has_garden = make_bernoulli('has_garden', N=1)
is_big = make_bernoulli('is_big', N=1)
is_green = cartesian_bernoulli_child('is_green', [is_big, has_garden], N=1)
model = pymc.Model([has_garden, is_big, is_green])
coeff_values = sample_coeffs(model)
# define identical model again but fix coefficient values
has_garden = make_bernoulli('has_garden', N=1, fixed=coeff_values)
is_big = make_bernoulli('is_big', N=1, fixed=coeff_values)
is_green = cartesian_bernoulli_child('is_green', [is_big, has_garden],
N=1,
fixed=coeff_values)
fx_model = pymc.Model([has_garden, is_big, is_green])
# sample from the model with fixed coefficients
fx_sampler = pymc.MCMC(fx_model)
fx_sampler.sample(iter=2000)
has_garden_sample = fx_sampler.trace('has_garden')[:]
is_big_sample = fx_sampler.trace('is_big')[:]
is_green_sample = fx_sampler.trace('is_green')[:]
# define identical model again but fix coefficient values
has_garden, cfs1 = make_bernoulli('has_garden',
value=has_garden_sample,
return_coeffs=True)
is_big, cfs2 = make_bernoulli('is_big',
value=is_big_sample,
return_coeffs=True)
is_green, cfs3 = cartesian_bernoulli_child('is_green',
[is_big, has_garden],
value=is_green_sample,
return_coeffs=True)
model = pymc.Model(cfs1 + cfs2 + cfs3)
sampler = pymc.MCMC(model)
sampler.sample(iter=2000, burn=1000)
for pymc_var in sampler.stochastics:
name = str(pymc_var)
mean_posterior = sampler.trace(name)[:].mean()
actual = coeff_values[name]
print "%s %.3f %.3f" % (pad(name, 30), mean_posterior, actual)
assert np.isclose(mean_posterior, actual, rtol=0.1, atol=0.1)
def get_rsw_data(size):
"""generates a data from the famous 'rain - sprinkler - wet grass' bayesian network
returns as a dataframe.
"""
coefficients = {
'CF: p(rain)': 0.2,
'CF: p(grass_wet | sprinkler=0 rain=0)': 0.,
'CF: p(grass_wet | sprinkler=0 rain=1)': 0.8,
'CF: p(grass_wet | sprinkler=1 rain=0)': 0.9,
'CF: p(grass_wet | sprinkler=1 rain=1)': 0.99,
'CF: p(sprinkler | rain=0)': 0.4,
'CF: p(sprinkler | rain=1)': 0.01
}
# define the bayesian net
rain = make_bernoulli('rain', N=1, fixed=coefficients)
sprinkler = cartesian_child('sprinkler',
parents=[rain],
levels=2,
N=1,
fixed=coefficients)
grass_wet = cartesian_child('grass_wet',
parents=[sprinkler, rain],
levels=2,
N=1,
fixed=coefficients)
# sample from the prior
model = pymc.Model([rain, sprinkler, grass_wet])
sampler = pymc.MCMC(model)
sampler.sample(iter=size, burn=0)
data = pd.DataFrame({
str(node): sampler.trace(str(node))[:].ravel() + 0
for node in sampler.stochastics
})
return data
def construct_net(self, df):
mileage = make_bernoulli('mileage', value=df.mileage)
age = make_bernoulli('age', value=df.age)
socio_econ = cartesian_child('socio_econ',
parents=[age],
value=df.socio_econ)
risk_aversion = cartesian_child('risk_aversion',
parents=[age, socio_econ],
value=df.risk_aversion)
senior_train = cartesian_child('senior_train',
parents=[age, risk_aversion],
value=df.senior_train)
good_student = cartesian_child('good_student',
parents=[age, socio_econ],
value=df.good_student)
extra_car = cartesian_child('extra_car',
parents=[socio_econ],
value=df.extra_car)
driving_skill = cartesian_child('driving_skill',
parents=[age, senior_train],
value=df.driving_skill)
driving_hist = cartesian_child('driving_hist',
parents=[driving_skill, risk_aversion],
value=df.driving_hist)
driving_quality = cartesian_child(
'driving_quality',
parents=[driving_skill, risk_aversion],
value=df.driving_quality)
make_model = cartesian_child('make_model',
parents=[risk_aversion, socio_econ],
value=df.make_model)
vehicle_year = cartesian_child('vehicle_year',
parents=[socio_econ],
value=df.vehicle_year)
antilock = cartesian_child('antilock',
parents=[make_model, vehicle_year],
value=df.antilock)
ruggedness = cartesian_child('ruggedness',
parents=[make_model, antilock],
value=df.ruggedness)
accident = cartesian_child(
'accident',
parents=[driving_quality, antilock, mileage],
value=df.accident)
airbag = cartesian_child('airbag',
parents=[make_model, vehicle_year],
value=df.airbag)
car_value = cartesian_child(
'car_value',
parents=[make_model, mileage, vehicle_year],
value=df.car_value)
home_base = cartesian_child('home_base',
parents=[risk_aversion, socio_econ],
value=df.home_base)
anti_theft = cartesian_child('anti_theft',
parents=[risk_aversion, socio_econ],
value=df.anti_theft)
theft = cartesian_child('theft',
parents=[car_value, home_base, anti_theft],
value=df.theft)
own_damage = cartesian_child('own_damage',
parents=[ruggedness, accident],
value=df.own_damage)
own_car_cost = cartesian_child('own_car_cost',
parents=[own_damage, car_value, theft],
value=df.own_car_cost)
cushioning = cartesian_child('cushioning',
parents=[ruggedness, accident],
value=df.cushioning)
medical_cost = cartesian_child('medical_cost',
parents=[age, cushioning, accident],
value=df.medical_cost)
liability_cost = cartesian_child('liability_cost',
parents=[accident],
value=df.liability_cost)
other_car_cost = cartesian_child('other_car_cost',
parents=[ruggedness, accident],
value=df.other_car_cost)
property_cost = cartesian_child('property_cost',
parents=[other_car_cost, own_car_cost],
value=df.property_cost)
model = pymc.Model([
mileage, age, socio_econ, good_student, extra_car, driving_skill,
risk_aversion, senior_train, driving_hist, driving_quality,
make_model, vehicle_year, antilock, ruggedness, accident, airbag,
car_value, home_base, anti_theft, theft, own_damage, own_car_cost,
cushioning, medical_cost, liability_cost, other_car_cost,
property_cost
])
return model
def get_car_data_random_coeffs(size):
mileage = make_bernoulli('mileage', N=1)
age = make_bernoulli('age', N=1)
socio_econ = cartesian_child('socio_econ', parents=[age], N=1)
risk_aversion = cartesian_child('risk_aversion',
parents=[age, socio_econ],
N=1)
senior_train = cartesian_child('senior_train',
parents=[age, risk_aversion],
N=1)
good_student = cartesian_child('good_student',
parents=[age, socio_econ],
N=1)
extra_car = cartesian_child('extra_car', parents=[socio_econ], N=1)
driving_skill = cartesian_child('driving_skill',
parents=[age, senior_train],
N=1)
driving_hist = cartesian_child('driving_hist',
parents=[driving_skill, risk_aversion],
N=1)
driving_quality = cartesian_child('driving_quality',
parents=[driving_skill, risk_aversion],
N=1)
make_model = cartesian_child('make_model',
parents=[risk_aversion, socio_econ],
N=1)
vehicle_year = cartesian_child('vehicle_year', parents=[socio_econ], N=1)
antilock = cartesian_child('antilock',
parents=[make_model, vehicle_year],
N=1)
ruggedness = cartesian_child('ruggedness',
parents=[make_model, antilock],
N=1)
accident = cartesian_child('accident',
parents=[driving_quality, antilock, mileage],
N=1)
airbag = cartesian_child('airbag', parents=[make_model, vehicle_year], N=1)
car_value = cartesian_child('car_value',
parents=[make_model, mileage, vehicle_year],
N=1)
home_base = cartesian_child('home_base',
parents=[risk_aversion, socio_econ],
N=1)
anti_theft = cartesian_child('anti_theft',
parents=[risk_aversion, socio_econ],
N=1)
theft = cartesian_child('theft',
parents=[car_value, home_base, anti_theft],
N=1)
own_damage = cartesian_child('own_damage',
parents=[ruggedness, accident],
N=1)
own_car_cost = cartesian_child('own_car_cost',
parents=[own_damage, car_value, theft],
N=1)
cushioning = cartesian_child('cushioning',
parents=[ruggedness, accident],
N=1)
medical_cost = cartesian_child('medical_cost',
parents=[age, cushioning, accident],
N=1)
liability_cost = cartesian_child('liability_cost', parents=[accident], N=1)
other_car_cost = cartesian_child('other_car_cost',
parents=[ruggedness, accident],
N=1)
property_cost = cartesian_child('property_cost',
parents=[other_car_cost, own_car_cost],
N=1)
model = pymc.Model([
mileage, age, socio_econ, good_student, extra_car, driving_skill,
risk_aversion, senior_train, driving_hist, driving_quality, make_model,
vehicle_year, antilock, ruggedness, accident, airbag, car_value,
home_base, anti_theft, theft, own_damage, own_car_cost, cushioning,
medical_cost, liability_cost, other_car_cost, property_cost
])
coefficients = sample_coeffs(model)
mileage = make_bernoulli('mileage', N=1, fixed=coefficients)
age = make_bernoulli('age', N=1, fixed=coefficients)
socio_econ = cartesian_child('socio_econ',
parents=[age],
N=1,
fixed=coefficients)
risk_aversion = cartesian_child('risk_aversion',
parents=[age, socio_econ],
N=1,
fixed=coefficients)
senior_train = cartesian_child('senior_train',
parents=[age, risk_aversion],
N=1,
fixed=coefficients)
good_student = cartesian_child('good_student',
parents=[age, socio_econ],
N=1,
fixed=coefficients)
extra_car = cartesian_child('extra_car',
parents=[socio_econ],
N=1,
fixed=coefficients)
driving_skill = cartesian_child('driving_skill',
parents=[age, senior_train],
N=1)
driving_hist = cartesian_child('driving_hist',
parents=[driving_skill, risk_aversion],
N=1,
fixed=coefficients)
driving_quality = cartesian_child('driving_quality',
parents=[driving_skill, risk_aversion],
N=1,
fixed=coefficients)
make_model = cartesian_child('make_model',
parents=[risk_aversion, socio_econ],
N=1,
fixed=coefficients)
vehicle_year = cartesian_child('vehicle_year',
parents=[socio_econ],
N=1,
fixed=coefficients)
antilock = cartesian_child('antilock',
parents=[make_model, vehicle_year],
N=1,
fixed=coefficients)
ruggedness = cartesian_child('ruggedness',
parents=[make_model, antilock],
N=1,
fixed=coefficients)
accident = cartesian_child('accident',
parents=[driving_quality, antilock, mileage],
N=1,
fixed=coefficients)
airbag = cartesian_child('airbag',
parents=[make_model, vehicle_year],
N=1,
fixed=coefficients)
car_value = cartesian_child('car_value',
parents=[make_model, mileage, vehicle_year],
N=1,
fixed=coefficients)
home_base = cartesian_child('home_base',
parents=[risk_aversion, socio_econ],
N=1,
fixed=coefficients)
anti_theft = cartesian_child('anti_theft',
parents=[risk_aversion, socio_econ],
N=1,
fixed=coefficients)
theft = cartesian_child('theft',
parents=[car_value, home_base, anti_theft],
N=1,
fixed=coefficients)
own_damage = cartesian_child('own_damage',
parents=[ruggedness, accident],
N=1,
fixed=coefficients)
own_car_cost = cartesian_child('own_car_cost',
parents=[own_damage, car_value, theft],
N=1,
fixed=coefficients)
cushioning = cartesian_child('cushioning',
parents=[ruggedness, accident],
N=1,
fixed=coefficients)
medical_cost = cartesian_child('medical_cost',
parents=[age, cushioning, accident],
N=1,
fixed=coefficients)
liability_cost = cartesian_child('liability_cost',
parents=[accident],
N=1,
fixed=coefficients)
other_car_cost = cartesian_child('other_car_cost',
parents=[ruggedness, accident],
N=1,
fixed=coefficients)
property_cost = cartesian_child('property_cost',
parents=[other_car_cost, own_car_cost],
N=1,
fixed=coefficients)
model = pymc.Model([
mileage, age, socio_econ, good_student, extra_car, driving_skill,
risk_aversion, senior_train, driving_hist, driving_quality, make_model,
vehicle_year, antilock, ruggedness, accident, airbag, car_value,
home_base, anti_theft, theft, own_damage, own_car_cost, cushioning,
medical_cost, liability_cost, other_car_cost, property_cost
])
sampler = pymc.MCMC(model)
sampler.sample(iter=size, burn=0)
data = pd.DataFrame({
str(node): sampler.trace(str(node))[:].ravel() + 0
for node in sampler.stochastics if not str(node).startswith('CF: ')
})
return data
def get_cancer_data_random_coeffs(size):
anxiety = make_bernoulli('anxiety', N=1)
peer_pressure = make_bernoulli('peer_pressure', N=1)
smoking = cartesian_child('smoking', parents=[anxiety, peer_pressure], N=1)
yellow_fingers = cartesian_child('yellow_fingers', parents=[smoking], N=1)
genetics = make_bernoulli('genetics', N=1)
lung_cancer = cartesian_child('lung_cancer',
parents=[smoking, genetics],
N=1)
allergy = make_bernoulli('allergy', N=1)
coughing = cartesian_child('coughing', parents=[allergy, lung_cancer], N=1)
fatigue = cartesian_child('fatigue', parents=[lung_cancer, coughing], N=1)
attention_disorder = cartesian_child('attention_disorder',
parents=[genetics],
N=1)
car_accident = cartesian_child('car_accident',
parents=[fatigue, attention_disorder],
N=1)
model = pymc.Model([
anxiety, peer_pressure, smoking, yellow_fingers, genetics, lung_cancer,
allergy, coughing, fatigue, attention_disorder, car_accident
])
coefficients = sample_coeffs(model)
anxiety = make_bernoulli('anxiety', N=1, fixed=coefficients)
peer_pressure = make_bernoulli('peer_pressure', N=1, fixed=coefficients)
smoking = cartesian_child('smoking',
parents=[anxiety, peer_pressure],
N=1,
fixed=coefficients)
yellow_fingers = cartesian_child('yellow_fingers',
parents=[smoking],
N=1,
fixed=coefficients)
genetics = make_bernoulli('genetics', N=1, fixed=coefficients)
lung_cancer = cartesian_child('lung_cancer',
parents=[smoking, genetics],
N=1,
fixed=coefficients)
allergy = make_bernoulli('allergy', N=1, fixed=coefficients)
coughing = cartesian_child('coughing',
parents=[allergy, lung_cancer],
N=1,
fixed=coefficients)
fatigue = cartesian_child('fatigue',
parents=[lung_cancer, coughing],
N=1,
fixed=coefficients)
attention_disorder = cartesian_child('attention_disorder',
parents=[genetics],
N=1,
fixed=coefficients)
car_accident = cartesian_child('car_accident',
parents=[fatigue, attention_disorder],
N=1,
fixed=coefficients)
# sample from the prior
model = pymc.Model([
anxiety, peer_pressure, smoking, yellow_fingers, genetics, lung_cancer,
allergy, coughing, fatigue, attention_disorder, car_accident
])
sampler = pymc.MCMC(model)
sampler.sample(iter=size, burn=0)
data = pd.DataFrame({
str(node): sampler.trace(str(node))[:].ravel() + 0
for node in sampler.stochastics
})
return data