def __init__(self, mean=0, sigma=1, name="", model=None):
super().__init__(name, model)
self.Var("v1", Normal.dist(mu=mean, sigma=sigma))
Normal("v2", mu=mean, sigma=sigma)
Normal("v3", mu=mean, sigma=HalfCauchy("sd", beta=10, testval=1.0))
Deterministic("v3_sq", self.v3**2)
Potential("p1", tt.constant(1))
def test_density_scaling_with_genarator(self):
# We have different size generators
def gen1():
i = 0
while True:
yield np.ones((10, 100)) * i
i += 1
def gen2():
i = 0
while True:
yield np.ones((20, 100)) * i
i += 1
# We have same size models
with pm.Model() as model1:
Normal('n', observed=gen1(), total_size=100)
p1 = theano.function([], model1.logpt)
with pm.Model() as model2:
gen_var = generator(gen2())
Normal('n', observed=gen_var, total_size=100)
p2 = theano.function([], model2.logpt)
# We want densities to be equal
for _ in range(10):
np.testing.assert_almost_equal(p1(), p2())
def test_acceptance_rate_against_coarseness(self):
"""Test that the acceptance rate increases
when the coarse model is closer to
the fine model."""
with Model() as coarse_model_0:
Normal("x", 5.0, 1.0)
with Model() as coarse_model_1:
Normal("x", 6.0, 2.0)
with Model() as coarse_model_2:
Normal("x", 20.0, 5.0)
possible_coarse_models = [
coarse_model_0, coarse_model_1, coarse_model_2
]
acc = []
with Model():
Normal("x", 5.0, 1.0)
for coarse_model in possible_coarse_models:
step = MLDA(coarse_models=[coarse_model],
subsampling_rates=3,
tune=True)
trace = sample(chains=1, draws=500, tune=100, step=step)
acc.append(trace.get_sampler_stats("accepted").mean())
assert acc[0] > acc[1] > acc[2], ("Acceptance rate is not "
"strictly increasing when"
"coarse model is closer to "
"fine model. Acceptance rates"
"were: {}".format(acc))
def __init__(self, mean=0, sd=1, name='', model=None):
super(DocstringModel, self).__init__(name, model)
self.Var('v1', Normal.dist(mu=mean, sd=sd))
Normal('v2', mu=mean, sd=sd)
Normal('v3', mu=mean, sd=HalfCauchy('sd', beta=10, testval=1.))
Deterministic('v3_sq', self.v3**2)
Potential('p1', tt.constant(1))
def test_density_scaling_with_genarator(self):
# We have different size generators
def true_dens():
g = gen1()
for i, point in enumerate(g):
yield stats.norm.logpdf(point).sum() * 10
t = true_dens()
# We have same size models
with pm.Model() as model1:
Normal('n', observed=gen1(), total_size=100)
p1 = theano.function([], model1.logpt)
with pm.Model() as model2:
gen_var = generator(gen2())
Normal('n', observed=gen_var, total_size=100)
p2 = theano.function([], model2.logpt)
for i in range(10):
_1, _2, _t = p1(), p2(), next(t)
np.testing.assert_almost_equal(_1,
_t,
decimal=select_by_precision(
float64=7,
float32=2)) # Value O(-50,000)
np.testing.assert_almost_equal(_1, _2)
def test_multiple_subsampling_rates(self):
"""Test that when you give a signle integer it is applied to all levels and
when you give a list the list is applied correctly."""
with Model() as coarse_model_0:
Normal("n", 0, 2.2, shape=(3, ))
with Model() as coarse_model_1:
Normal("n", 0, 2.1, shape=(3, ))
with Model():
Normal("n", 0, 2.0, shape=(3, ))
step_1 = MLDA(coarse_models=[coarse_model_0, coarse_model_1],
subsampling_rates=3)
assert len(step_1.subsampling_rates) == 2
assert step_1.subsampling_rates[0] == step_1.subsampling_rates[
1] == 3
step_2 = MLDA(coarse_models=[coarse_model_0, coarse_model_1],
subsampling_rates=[3, 4])
assert step_2.subsampling_rates[0] == 3
assert step_2.subsampling_rates[1] == 4
with pytest.raises(ValueError):
step_3 = MLDA(
coarse_models=[coarse_model_0, coarse_model_1],
subsampling_rates=[3, 4, 10],
)
def test_density_scaling(self):
with pm.Model() as model1:
Normal('n', observed=[[1]], total_size=1)
p1 = theano.function([], model1.logpt)
with pm.Model() as model2:
Normal('n', observed=[[1]], total_size=2)
p2 = theano.function([], model2.logpt)
self.assertEqual(p1() * 2, p2())
def test_float64(self):
with Model() as model:
x = Normal('x', testval=np.array(1., dtype='float64'))
obs = Normal('obs', mu=x, sigma=1., observed=np.random.randn(5))
assert x.dtype == 'float64'
assert obs.dtype == 'float64'
for sampler in self.samplers:
with model:
sample(10, sampler())
def test_float64(self):
with Model() as model:
x = Normal("x", testval=np.array(1.0, dtype="float64"))
obs = Normal("obs", mu=x, sigma=1.0, observed=np.random.randn(5))
assert x.dtype == "float64"
assert obs.dtype == "float64"
for sampler in self.samplers:
with model:
sample(10, sampler())
def check_trace(self, step_method):
"""Tests whether the trace for step methods is exactly the same as on master.
Code changes that effect how random numbers are drawn may change this, and require
`master_samples` to be updated, but such changes should be noted and justified in the
commit.
This method may also be used to benchmark step methods across commits, by running, for
example
```
BENCHMARK=100000 ./scripts/test.sh -s pymc3/tests/test_step.py:TestStepMethods
```
on multiple commits.
"""
n_steps = 100
with Model() as model:
x = Normal("x", mu=0, sd=1)
y = Normal("y", mu=x, sd=1, observed=1)
if step_method.__name__ == "SMC":
trace = sample(draws=200,
random_seed=1,
progressbar=False,
step=step_method())
elif step_method.__name__ == "NUTS":
step = step_method(scaling=model.test_point)
trace = sample(
0,
tune=n_steps,
discard_tuned_samples=False,
step=step,
random_seed=1,
chains=1,
)
else:
trace = sample(
0,
tune=n_steps,
discard_tuned_samples=False,
step=step_method(),
random_seed=1,
chains=1,
)
assert_array_almost_equal(
trace["x"],
self.master_samples[step_method],
decimal=select_by_precision(float64=6, float32=4),
)
def test_float64(self):
theano.config.floatX = 'float64'
theano.config.warn_float64 = 'ignore'
with Model() as model:
x = Normal('x', testval=np.array(1., dtype='float64'))
obs = Normal('obs', mu=x, sd=1., observed=np.random.randn(5))
assert x.dtype == 'float64'
assert obs.dtype == 'float64'
for sampler in self.samplers:
with model:
sample(10, sampler())
def test_linalg(self):
with Model():
a = Normal('a', shape=2)
a = tt.switch(a > 0, np.inf, a)
b = tt.slinalg.solve(floatX(np.eye(2)), a)
Normal('c', mu=b, shape=2)
with warnings.catch_warnings(record=True) as warns:
trace = sample(20, init=None, tune=5)
assert np.any(trace['diverging'])
assert any('diverging samples after tuning' in str(warn.message)
for warn in warns)
assert any('contains only' in str(warn.message) for warn in warns)
with pytest.raises(SamplingError):
sample(20, init=None, nuts_kwargs={'on_error': 'raise'})
def test_trace_length(self):
"""Check if trace length is as expected."""
tune = 100
draws = 50
with Model() as coarse_model:
Normal("n", 0, 2.2, shape=(3, ))
with Model():
Normal("n", 0, 2, shape=(3, ))
step = MLDA(coarse_models=[coarse_model])
trace = sample(tune=tune,
draws=draws,
step=step,
chains=1,
discard_tuned_samples=False)
assert len(trace) == tune + draws
def test_constant_step():
with Model() as model:
x = Normal('x', 0, 1)
start = {'x': -1}
tr = sample(10, step=Constant([x]), start=start)
assert_almost_equal(tr['x'], start['x'], decimal=10)
def test_competence(self, variable, has_grad, outcome):
with Model() as pmodel:
Normal('n', 0, 2, shape=(3, ))
Binomial('b', n=2, p=0.3)
assert DEMetropolisZ.competence(pmodel[variable],
has_grad=has_grad) == outcome
pass
def test_competence(self, variable, has_grad, outcome):
"""Test if competence function returns expected
results for different models"""
with Model() as pmodel:
Normal("n", 0, 2, shape=(3, ))
Binomial("b", n=2, p=0.3)
assert MLDA.competence(pmodel[variable], has_grad=has_grad) == outcome
def test_parallelized_chains_are_random(self):
"""Test that parallel chain are
not identical when parallelisation
is applied"""
with Model() as coarse_model:
Normal("x", 0.3, 1)
with Model():
Normal("x", 0, 1)
for stepper in TestMLDA.steppers:
step = stepper(coarse_models=[coarse_model])
trace = sample(chains=2, cores=2, draws=20, tune=0, step=step)
samples = np.array(trace.get_values("x", combine=False))[:, 5]
assert (
len(set(samples)) == 2
), "Parallelized {} " "chains are identical.".format(stepper)
def __init__(self, mean=0, sigma=1, name='', model=None):
super().__init__(name, model)
self.Var('v1', Normal.dist(mu=mean, sigma=sigma))
Normal('v2', mu=mean, sigma=sigma)
Normal('v3', mu=mean, sigma=HalfCauchy('sd', beta=10, testval=1.))
Deterministic('v3_sq', self.v3 ** 2)
Potential('p1', tt.constant(1))
def test_assign_step_methods():
with Model() as model:
x = Bernoulli('x', 0.5)
steps = assign_step_methods(model, [])
assert isinstance(steps, BinaryMetropolis)
with Model() as model:
x = Normal('x', 0, 1)
steps = assign_step_methods(model, [])
assert isinstance(steps, NUTS)
with Model() as model:
x = Categorical('x', np.array([0.25, 0.75]))
steps = assign_step_methods(model, [])
assert isinstance(steps, BinaryMetropolis)
with Model() as model:
x = Categorical('x', np.array([0.25, 0.70, 0.05]))
steps = assign_step_methods(model, [])
assert isinstance(steps, Metropolis)
with Model() as model:
x = Binomial('x', 10, 0.5)
steps = assign_step_methods(model, [])
assert isinstance(steps, Metropolis)
def check_trace(self, step_method):
"""Tests whether the trace for step methods is exactly the same as on master.
Code changes that effect how random numbers are drawn may change this, and require
`master_samples` to be updated, but such changes should be noted and justified in the
commit.
This method may also be used to benchmark step methods across commits, by running, for
example
```
BENCHMARK=100000 ./scripts/test.sh -s pymc3/tests/test_step.py:TestStepMethods
```
on multiple commits.
"""
test_steps = 100
n_steps = int(os.getenv('BENCHMARK', 100))
benchmarking = (n_steps != test_steps)
if benchmarking:
tqdm.write('Benchmarking {} with {:,d} samples'.format(step_method.__name__, n_steps))
else:
tqdm.write('Checking {} has same trace as on master'.format(step_method.__name__))
with Model() as model:
Normal('x', mu=0, sd=1)
trace = sample(n_steps, step=step_method(), random_seed=1)
if not benchmarking:
assert_array_almost_equal(trace.get_values('x'), self.master_samples[step_method])
def test_sampler_stats(self):
with Model() as model:
x = Normal("x", mu=0, sigma=1)
trace = sample(draws=10, tune=1, chains=1)
# Assert stats exist and have the correct shape.
expected_stat_names = {
"depth",
"diverging",
"energy",
"energy_error",
"model_logp",
"max_energy_error",
"mean_tree_accept",
"step_size",
"step_size_bar",
"tree_size",
"tune",
}
assert trace.stat_names == expected_stat_names
for varname in trace.stat_names:
assert trace.get_sampler_stats(varname).shape == (10, )
# Assert model logp is computed correctly: computing post-sampling
# and tracking while sampling should give same results.
model_logp_ = np.array([
model.logp(trace.point(i, chain=c)) for c in trace.chains
for i in range(len(trace))
])
assert (trace.model_logp == model_logp_).all()
def check_trace(self, step_method):
"""Tests whether the trace for step methods is exactly the same as on master.
Code changes that effect how random numbers are drawn may change this, and require
`master_samples` to be updated, but such changes should be noted and justified in the
commit.
This method may also be used to benchmark step methods across commits, by running, for
example
```
BENCHMARK=100000 ./scripts/test.sh -s pymc3/tests/test_step.py:TestStepMethods
```
on multiple commits.
"""
n_steps = 100
with Model():
x = Normal('x', mu=0, sd=1)
if step_method.__name__ == 'SMC':
Deterministic('like', - 0.5 * tt.log(2 * np.pi) - 0.5 * x.T.dot(x))
trace = smc.ATMIP_sample(n_steps=n_steps, step=step_method(random_seed=1),
n_jobs=1, progressbar=False,
homepath=self.temp_dir)
else:
trace = sample(0, tune=n_steps,
discard_tuned_samples=False,
step=step_method(), random_seed=1)
assert_array_almost_equal(
trace.get_values('x'),
self.master_samples[step_method],
decimal=select_by_precision(float64=6, float32=4))
def test_float32_MLDA(self):
data = np.random.randn(5).astype("float32")
with Model() as coarse_model:
x = Normal("x", testval=np.array(1.0, dtype="float32"))
obs = Normal("obs", mu=x, sigma=1.0, observed=data + 0.5)
with Model() as model:
x = Normal("x", testval=np.array(1.0, dtype="float32"))
obs = Normal("obs", mu=x, sigma=1.0, observed=data)
assert x.dtype == "float32"
assert obs.dtype == "float32"
with model:
sample(10, MLDA(coarse_models=[coarse_model]))
def test_float32_MLDA(self):
data = np.random.randn(5).astype('float32')
with Model() as coarse_model:
x = Normal('x', testval=np.array(1., dtype='float32'))
obs = Normal('obs', mu=x, sigma=1., observed=data + 0.5)
with Model() as model:
x = Normal('x', testval=np.array(1., dtype='float32'))
obs = Normal('obs', mu=x, sigma=1., observed=data)
assert x.dtype == 'float32'
assert obs.dtype == 'float32'
with model:
sample(10, MLDA(coarse_models=[coarse_model]))
def test_gradient_with_scaling(self):
with pm.Model() as model1:
genvar = generator(gen1())
m = Normal('m')
Normal('n', observed=genvar, total_size=1000)
grad1 = theano.function([m], tt.grad(model1.logpt, m))
with pm.Model() as model2:
m = Normal('m')
shavar = theano.shared(np.ones((1000, 100)))
Normal('n', observed=shavar)
grad2 = theano.function([m], tt.grad(model2.logpt, m))
for i in range(10):
shavar.set_value(np.ones((100, 100)) * i)
g1 = grad1(1)
g2 = grad2(1)
np.testing.assert_almost_equal(g1, g2)
def test_posterior_estimate(self):
alpha_true, sigma_true = 1., 0.5
beta_true = 1.
size = 1000
X = np.random.randn(size)
Y = alpha_true + beta_true * X + np.random.randn(size) * sigma_true
decimal = 1
with Model() as model:
alpha = Normal('alpha', mu=0, sd=100, testval=alpha_true)
beta = Normal('beta', mu=0, sd=100, testval=beta_true)
sigma = InverseGamma('sigma', 10., testval=sigma_true)
mu = alpha + beta * X
Y_obs = Normal('Y_obs', mu=mu, sd=sigma, observed=Y)
for step_method, params in ((NUTS, {
"target_accept": 0.95
}), (Slice, {}), (Metropolis, {
'scaling': 10.
})):
trace = sample(100000,
step=step_method(**params),
progressbar=False,
tune=1000)
trace_ = trace[-300::5]
# We do the same for beta - using more burnin.
np.testing.assert_almost_equal(np.mean(trace_.alpha),
alpha_true,
decimal=decimal)
np.testing.assert_almost_equal(np.mean(trace_.beta),
beta_true,
decimal=decimal)
np.testing.assert_almost_equal(np.mean(trace_.sigma),
sigma_true,
decimal=decimal)
# Make sure posteriors are normal
_, p_alpha = stats.normaltest(trace_.alpha)
_, p_beta = stats.normaltest(trace_.beta)
# p-values should be > .05 to indiciate
np.testing.assert_array_less(0.05, p_alpha, verbose=True)
np.testing.assert_array_less(0.05, p_beta, verbose=True)
def test_linalg(self, caplog):
with Model():
a = Normal('a', shape=2)
a = tt.switch(a > 0, np.inf, a)
b = tt.slinalg.solve(floatX(np.eye(2)), a)
Normal('c', mu=b, shape=2)
caplog.clear()
trace = sample(20, init=None, tune=5, chains=2)
warns = [msg.msg for msg in caplog.records]
assert np.any(trace['diverging'])
assert (any('divergences after tuning' in warn for warn in warns)
or any('only diverging samples' in warn for warn in warns))
with pytest.raises(ValueError) as error:
trace.report.raise_ok()
error.match('issues during sampling')
assert not trace.report.ok
def test_linalg(self):
with Model():
a = Normal('a', shape=2)
a = tt.switch(a > 0, np.inf, a)
b = tt.slinalg.solve(floatX(np.eye(2)), a)
Normal('c', mu=b, shape=2)
with pytest.warns(None) as warns:
trace = sample(20, init=None, tune=5)
warns = [str(warn.message) for warn in warns]
assert np.any(trace['diverging'])
assert any('diverging samples after tuning' in warn
for warn in warns)
# FIXME This test fails sporadically on py27.
# It seems that capturing warnings doesn't work as expected.
# assert any('contains only' in warn for warn in warns)
with pytest.raises(SamplingError):
sample(20, init=None, nuts_kwargs={'on_error': 'raise'})
def test_normal_nograd_op(self):
"""Test normal distribution without an implemented gradient is assigned slice method"""
with Model() as model:
x = Normal('x', 0, 1)
# a custom Theano Op that does not have a grad:
is_64 = theano.config.floatX == "float64"
itypes = [tt.dscalar] if is_64 else [tt.fscalar]
otypes = [tt.dscalar] if is_64 else [tt.fscalar]
@theano.as_op(itypes, otypes)
def kill_grad(x):
return x
data = np.random.normal(size=(100,))
Normal("y", mu=kill_grad(x), sd=1, observed=data.astype(theano.config.floatX))
steps = assign_step_methods(model, [])
assert isinstance(steps, Slice)
def test_checks_population_size(self):
"""Test that population samplers check the population size."""
with Model() as model:
n = Normal('n', mu=0, sd=1)
for stepper in TestPopulationSamplers.steppers:
step = stepper()
with pytest.raises(ValueError):
trace = sample(draws=100, chains=1, step=step)
trace = sample(draws=100, chains=4, step=step)
pass
def test_checks_population_size(self):
"""Test that population samplers check the population size."""
with Model() as model:
n = Normal("n", mu=0, sigma=1)
for stepper in TestPopulationSamplers.steppers:
step = stepper()
with pytest.raises(ValueError):
sample(draws=10, tune=10, chains=1, cores=1, step=step)
# don't parallelize to make test faster
sample(draws=10, tune=10, chains=4, cores=1, step=step)
pass