def test_tag_output(self):
def model():
return ergo.normal(7, 1, name="x")
samples = ergo.run(model, num_samples=2000)
stats = samples.describe()
assert 6 < stats["output"]["mean"] < 8
def test_rejection():
def model():
x = ergo.flip()
y = ergo.flip()
ergo.condition(x or y)
return x == y
samples = ergo.run(model, num_samples=1000)
assert 266 < sum(samples["output"]) < 466
def test_mem_2():
"""
Check that mem is cleared at the start of each run
"""
@ergo.mem
def model():
return ergo.lognormal_from_interval(1, 10)
samples = ergo.run(model, num_samples=100)
assert samples["output"].unique().size == 100
def test_sampling(self):
def model():
x = ergo.lognormal_from_interval(1, 10, name="x")
y = ergo.beta_from_hits(1, 9, name="y")
z = x * y
ergo.tag(z, "z")
samples = ergo.run(model, num_samples=2000)
stats = samples.describe()
assert 3.5 < stats["x"]["mean"] < 4.5
assert 0.1 < stats["y"]["mean"] < 0.3
assert 0.6 < stats["z"]["mean"] < 1.0
def test_nomem():
"""
Without mem, different calls to foo() should differ sometimes
"""
def foo():
return ergo.lognormal_from_interval(1, 10)
def model():
x = foo()
y = foo()
return x == y
samples = ergo.run(model, num_samples=1000)
assert sum(samples["output"]) != 1000
def test_mem():
"""
With mem, different calls to foo() should always have the same value
"""
@ergo.mem
def foo():
return ergo.lognormal_from_interval(1, 10)
def model():
x = foo()
y = foo()
return x == y
samples = ergo.run(model, num_samples=1000)
assert sum(samples["output"]) == 1000
def test_foretold_sampling(self):
foretold = ergo.Foretold()
# https://www.foretold.io/c/f45577e4-f1b0-4bba-8cf6-63944e63d70c/m/cf86da3f-c257-4787-b526-3ef3cb670cb4
# Distribution is mm(10 to 20, 200 to 210), a mixture model with most mass split between
# 10 - 20 and 200 - 210.
dist = foretold.get_question("cf86da3f-c257-4787-b526-3ef3cb670cb4")
assert dist.quantile(0.25) < 100
assert dist.quantile(0.75) > 100
num_samples = 20000
samples = ergo.run(lambda: ergo.tag(dist.sample_community(), "sample"),
num_samples=num_samples)
# Probability mass is split evenly between both modes of the distribution, so approximately half of the
# samples should be lower than 100
assert np.count_nonzero(samples > 100) == pytest.approx(
num_samples / 2, 0.1)
def plot_question(sampler, num_samples=200, bw=None, start_date=date.today()):
samples = ergo.run(sampler, num_samples=num_samples)
summarize_question_samples(samples)
q = sampler.question
q_samples = samples[sampler.__name__]
if (
q.id == 4128
): # Date question: Need to convert back to date from days (https://github.com/oughtinc/ergo/issues/144)
q_samples = np.array([start_date + timedelta(s) for s in q_samples])
if bw is not None:
q.show_prediction(samples=q_samples,
show_community=True,
percent_kept=0.9,
bw=bw)
else:
q.show_prediction(samples=q_samples,
show_community=True,
percent_kept=0.9)