def test_multiple_samplers(self, caplog):
with Model():
prob = Beta("prob", alpha=5.0, beta=3.0)
Binomial("outcome", n=1, p=prob)
caplog.clear()
sample(3, tune=2, discard_tuned_samples=False, n_init=None, chains=1)
messages = [msg.msg for msg in caplog.records]
assert all("boolean index did not" not in msg for msg in messages)
def test_multiple_samplers(self):
with Model():
prob = Beta('prob', alpha=5., beta=3.)
Binomial('outcome', n=1, p=prob)
with pytest.warns(None) as warns:
sample(3, tune=2, discard_tuned_samples=False,
n_init=None)
messages = [warn.message.args[0] for warn in warns]
assert any("contains only 3" in msg for msg in messages)
assert all('boolean index did not' not in msg for msg in messages)
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):
trace = sample(draws=100, chains=1, step=step)
trace = sample(draws=100, chains=4, step=step)
pass
def test_float32(self):
with Model() as model:
x = Normal('x', testval=np.array(1., dtype='float32'))
obs = Normal('obs', mu=x, sigma=1., observed=np.random.randn(5).astype('float32'))
assert x.dtype == 'float32'
assert obs.dtype == 'float32'
for sampler in self.samplers:
with model:
sample(10, sampler())
def test_multiple_samplers(self):
with Model():
prob = Beta('prob', alpha=5., beta=3.)
Binomial('outcome', n=1, p=prob)
# Catching warnings through multiprocessing doesn't work,
# so we have to use single threaded sampling.
with pytest.warns(None) as warns:
sample(3, tune=2, discard_tuned_samples=False,
n_init=None, chains=1)
messages = [warn.message.args[0] for warn in warns]
assert any("contains only 3" in msg for msg in messages)
assert all('boolean index did not' not in msg for msg in messages)
def test_multiple_samplers(self, caplog):
with Model():
prob = Beta("prob", alpha=5.0, beta=3.0)
Binomial("outcome", n=1, p=prob)
caplog.clear()
sample(3,
tune=2,
discard_tuned_samples=False,
n_init=None,
chains=1)
messages = [msg.msg for msg in caplog.records]
assert all("boolean index did not" not in msg for msg in messages)
def test_checks_population_size(self):
"""Test that population samplers check the population size."""
steppers = [
DEMetropolis
]
with Model() as model:
n = Normal('n', mu=0, sd=1)
for stepper in 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 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)
if step_method.__name__ == 'SMC':
trace = smc.sample_smc(n_steps=n_steps,
n_chains=2,
start=[{
'x': 1.
}, {
'x': -1.
}],
random_seed=1,
n_jobs=1,
progressbar=False,
homepath=self.temp_dir)
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.get_values('x'),
self.master_samples[step_method],
decimal=select_by_precision(float64=6,
float32=4))
def test_float32(self):
with Model() as model:
x = Normal("x", testval=np.array(1.0, dtype="float32"))
obs = Normal("obs",
mu=x,
sigma=1.0,
observed=np.random.randn(5).astype("float32"))
assert x.dtype == "float32"
assert obs.dtype == "float32"
for sampler in self.samplers:
with model:
sample(10, sampler())
def test_float32(self):
with Model() as model:
x = Normal('x', testval=np.array(1., dtype='float32'))
obs = Normal('obs',
mu=x,
sigma=1.,
observed=np.random.randn(5).astype('float32'))
assert x.dtype == 'float32'
assert obs.dtype == 'float32'
for sampler in self.samplers:
with model:
sample(10, sampler())
def test_float32(self):
theano.config.floatX = 'float32'
theano.config.warn_float64 = 'warn'
with Model() as model:
x = Normal('x', testval=np.array(1., dtype='float32'))
obs = Normal('obs', mu=x, sd=1., observed=np.random.randn(5).astype('float32'))
assert x.dtype == 'float32'
assert obs.dtype == 'float32'
for sampler in self.samplers:
with model:
sample(10, sampler())
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 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_demcmc_warning_on_small_populations(self):
"""Test that a warning is raised when n_chains <= n_dims"""
with Model() as model:
Normal("n", mu=0, sigma=1, shape=(2, 3))
with pytest.warns(UserWarning) as record:
sample(
draws=5,
tune=5,
chains=6,
step=DEMetropolis(),
# make tests faster by not parallelizing; disable convergence warning
cores=1,
compute_convergence_checks=False)
pass
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, sigma=1)
y = Normal("y", mu=x, sigma=1, observed=1)
if step_method.__name__ == "SMC":
trace = sample(
draws=200, random_seed=1, progressbar=False, step=step_method(parallel=False)
)
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_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)
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_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_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 test_step_continuous():
start, model, (mu, C) = mv_simple()
with model:
mh = Metropolis()
slicer = Slice()
hmc = HamiltonianMC(scaling=C, is_cov=True, blocked=False)
nuts = NUTS(scaling=C, is_cov=True, blocked=False)
mh_blocked = Metropolis(S=C,
proposal_dist=MultivariateNormalProposal,
blocked=True)
slicer_blocked = Slice(blocked=True)
hmc_blocked = HamiltonianMC(scaling=C, is_cov=True)
nuts_blocked = NUTS(scaling=C, is_cov=True)
compound = CompoundStep([hmc_blocked, mh_blocked])
steps = [slicer, hmc, nuts, mh_blocked, hmc_blocked,
slicer_blocked, nuts_blocked, compound]
unc = np.diag(C) ** .5
check = [('x', np.mean, mu, unc / 10.),
('x', np.std, unc, unc / 10.)]
for st in steps:
h = sample(8000, st, start, model=model, random_seed=1)
for (var, stat, val, bound) in check:
yield check_stat, repr(st), h, var, stat, val, bound
def test_parallel_start():
model, _, _, _ = simple_init()
with model:
tr = sample(5, njobs=2, start=[{'x': [10, 10]}, {
'x': [-10, -10]}], random_seed=RSEED)
assert tr.get_values('x', chains=0)[0][0] > 0
assert tr.get_values('x', chains=1)[0][0] < 0
def test_linear():
lam = -0.78
sig2 = 5e-3
N = 300
dt = 1e-1
sde = lambda x, lam: (lam * x, sig2)
x = floatX(_gen_sde_path(sde, (lam, ), dt, N, 5.0))
z = x + np.random.randn(x.size) * sig2
# build model
with Model() as model:
lamh = Flat("lamh")
xh = EulerMaruyama("xh", dt, sde, (lamh, ), shape=N + 1, testval=x)
Normal("zh", mu=xh, sigma=sig2, observed=z)
# invert
with model:
trace = sample(init="advi+adapt_diag", chains=1)
ppc = sample_posterior_predictive(trace, model=model)
ppcf = fast_sample_posterior_predictive(trace, model=model)
# test
p95 = [2.5, 97.5]
lo, hi = np.percentile(trace[lamh], p95, axis=0)
assert (lo < lam) and (lam < hi)
lo, hi = np.percentile(ppc["zh"], p95, axis=0)
assert ((lo < z) * (z < hi)).mean() > 0.95
lo, hi = np.percentile(ppcf["zh"], p95, axis=0)
assert ((lo < z) * (z < hi)).mean() > 0.95
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_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 test_step_continuous(self):
start, model, (mu, C) = mv_simple()
unc = np.diag(C)**.5
check = (('x', np.mean, mu, unc / 10.), ('x', np.std, unc, unc / 10.))
with model:
steps = (
Slice(),
HamiltonianMC(scaling=C, is_cov=True, blocked=False),
NUTS(scaling=C, is_cov=True, blocked=False),
Metropolis(S=C,
proposal_dist=MultivariateNormalProposal,
blocked=True),
Slice(blocked=True),
HamiltonianMC(scaling=C, is_cov=True),
NUTS(scaling=C, is_cov=True),
CompoundStep([
HamiltonianMC(scaling=C, is_cov=True),
HamiltonianMC(scaling=C, is_cov=True, blocked=False)
]),
)
for step in steps:
trace = sample(8000,
step=step,
start=start,
model=model,
random_seed=1)
yield self.check_stat, check, trace, step.__class__.__name__
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 test_step_continuous():
start, model, (mu, C) = mv_simple()
with model:
mh = Metropolis()
slicer = Slice()
hmc = HamiltonianMC(scaling=C, is_cov=True, blocked=False)
nuts = NUTS(scaling=C, is_cov=True, blocked=False)
mh_blocked = Metropolis(S=C,
proposal_dist=MultivariateNormalProposal,
blocked=True)
slicer_blocked = Slice(blocked=True)
hmc_blocked = HamiltonianMC(scaling=C, is_cov=True)
nuts_blocked = NUTS(scaling=C, is_cov=True)
compound = CompoundStep([hmc_blocked, mh_blocked])
steps = [
slicer, hmc, nuts, mh_blocked, hmc_blocked, slicer_blocked,
nuts_blocked, compound
]
unc = np.diag(C)**.5
check = [('x', np.mean, mu, unc / 10.), ('x', np.std, unc, unc / 10.)]
for st in steps:
h = sample(8000, st, start, model=model, random_seed=1)
for (var, stat, val, bound) in check:
yield check_stat, repr(st), h, var, stat, val, bound
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_step_continuous(self):
start, model, (mu, C) = mv_simple()
unc = np.diag(C) ** 0.5
check = (("x", np.mean, mu, unc / 10.0), ("x", np.std, unc, unc / 10.0))
with model:
steps = (
Slice(),
HamiltonianMC(scaling=C, is_cov=True, blocked=False),
NUTS(scaling=C, is_cov=True, blocked=False),
Metropolis(S=C, proposal_dist=MultivariateNormalProposal, blocked=True),
Slice(blocked=True),
HamiltonianMC(scaling=C, is_cov=True),
NUTS(scaling=C, is_cov=True),
CompoundStep(
[
HamiltonianMC(scaling=C, is_cov=True),
HamiltonianMC(scaling=C, is_cov=True, blocked=False),
]
),
)
for step in steps:
trace = sample(
0,
tune=8000,
chains=1,
discard_tuned_samples=False,
step=step,
start=start,
model=model,
random_seed=1,
)
self.check_stat(check, trace, step.__class__.__name__)
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_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_tuning_and_scaling_on(self):
"""Test that tune and base_scaling change as expected when
tuning is on."""
np.random.seed(1234)
ts = 100
_, model = simple_2model_continuous()
_, model_coarse = simple_2model_continuous()
with model:
trace = sample(
tune=ts,
draws=20,
step=MLDA(
coarse_models=[model_coarse],
base_tune_interval=50,
base_scaling=100.0,
),
chains=1,
discard_tuned_samples=False,
random_seed=1234,
)
assert trace.get_sampler_stats("tune", chains=0)[0]
assert trace.get_sampler_stats("tune", chains=0)[ts - 1]
assert not trace.get_sampler_stats("tune", chains=0)[ts]
assert not trace.get_sampler_stats("tune", chains=0)[-1]
assert trace.get_sampler_stats("base_scaling", chains=0)[0][0] == 100.0
assert trace.get_sampler_stats("base_scaling", chains=0)[0][1] == 100.0
assert trace.get_sampler_stats("base_scaling", chains=0)[-1][0] < 100.0
assert trace.get_sampler_stats("base_scaling", chains=0)[-1][1] < 100.0
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_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 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():
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 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, stage='0',
homepath=self.temp_dir)
else:
trace = sample(n_steps, step=step_method(), random_seed=1)
print(repr(trace.get_values('x')))
assert_array_almost_equal(
trace.get_values('x'),
self.master_samples[step_method],
decimal=select_by_precision(float64=6, float32=4))
def test_step_continuous(self):
start, model, (mu, C) = mv_simple()
unc = np.diag(C)**0.5
check = (("x", np.mean, mu, unc / 10.0), ("x", np.std, unc,
unc / 10.0))
with model:
steps = (
Slice(),
HamiltonianMC(scaling=C, is_cov=True, blocked=False),
NUTS(scaling=C, is_cov=True, blocked=False),
Metropolis(S=C,
proposal_dist=MultivariateNormalProposal,
blocked=True),
Slice(blocked=True),
HamiltonianMC(scaling=C, is_cov=True),
NUTS(scaling=C, is_cov=True),
CompoundStep([
HamiltonianMC(scaling=C, is_cov=True),
HamiltonianMC(scaling=C, is_cov=True, blocked=False),
]),
)
for step in steps:
trace = sample(
0,
tune=8000,
chains=1,
discard_tuned_samples=False,
step=step,
start=start,
model=model,
#random_seed=1,
cores=1)
self.check_stat(check, trace, step.__class__.__name__)
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_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]
print(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_step_elliptical_slice(self):
start, model, (K, L, mu, std, noise) = mv_prior_simple()
unc = noise ** 0.5
check = (("x", np.mean, mu, unc / 10.0), ("x", np.std, std, unc / 10.0))
with model:
steps = (EllipticalSlice(prior_cov=K), EllipticalSlice(prior_chol=L))
for step in steps:
trace = sample(
5000, tune=0, step=step, start=start, model=model, random_seed=1, chains=1
)
self.check_stat(check, trace, step.__class__.__name__)
def test_custom_proposal_dist(self):
with Model() as pmodel:
D = 3
Normal('n', 0, 2, shape=(D, ))
trace = sample(tune=100,
draws=50,
step=DEMetropolisZ(proposal_dist=NormalProposal),
cores=1,
chains=3,
discard_tuned_samples=False)
pass
def test_step_discrete(self):
start, model, (mu, C) = mv_simple_discrete()
unc = np.diag(C) ** .5
check = (('x', np.mean, mu, unc / 10.),
('x', np.std, unc, unc / 10.))
with model:
steps = (
Metropolis(S=C, proposal_dist=MultivariateNormalProposal),
)
for step in steps:
trace = sample(20000, step=step, start=start, model=model, random_seed=1)
yield self.check_stat, check, trace, step.__class__.__name__
def test_step_categorical(self):
start, model, (mu, C) = simple_categorical()
unc = C ** 0.5
check = (("x", np.mean, mu, unc / 10.0), ("x", np.std, unc, unc / 10.0))
with model:
steps = (
CategoricalGibbsMetropolis(model.x, proposal="uniform"),
CategoricalGibbsMetropolis(model.x, proposal="proportional"),
)
for step in steps:
trace = sample(8000, tune=0, step=step, start=start, model=model, random_seed=1)
self.check_stat(check, trace, step.__class__.__name__)
def test_step_discrete(self):
start, model, (mu, C) = mv_simple_discrete()
unc = np.diag(C) ** .5
check = (('x', np.mean, mu, unc / 10.),
('x', np.std, unc, unc / 10.))
with model:
steps = (
Metropolis(S=C, proposal_dist=MultivariateNormalProposal),
)
for step in steps:
trace = sample(20000, step=step, start=start, model=model, random_seed=1)
self.check_stat(check, trace)
def test_parallelized_chains_are_random(self):
with Model() as model:
x = Normal("x", 0, 1)
for stepper in TestPopulationSamplers.steppers:
step = stepper()
trace = sample(chains=4, cores=4, draws=20, tune=0, step=DEMetropolis())
samples = np.array(trace.get_values("x", combine=False))[:, 5]
assert len(set(samples)) == 4, "Parallelized {} " "chains are identical.".format(
stepper
)
pass
def test_step_elliptical_slice(self):
start, model, (K, mu, noise) = mv_prior_simple()
unc = noise ** 0.5
check = (('x', np.mean, mu, unc / 10.),
('x', np.std, unc, unc / 10.))
with model:
steps = (
EllipticalSlice(prior_cov=K),
)
for step in steps:
trace = sample(8000, step=step, start=start, model=model, random_seed=1)
yield self.check_stat, check, trace, step.__class__.__name__
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)
if step_method.__name__ == 'SMC':
trace = smc.sample_smc(n_steps=n_steps,
n_chains=2,
start=[{'x':1.}, {'x':-1.}],
random_seed=1,
n_jobs=1, progressbar=False,
homepath=self.temp_dir)
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.get_values('x'),
self.master_samples[step_method],
decimal=select_by_precision(float64=6, float32=4))
def test_interval_missing_observations():
with Model() as model:
obs1 = ma.masked_values([1, 2, -1, 4, -1], value=-1)
obs2 = ma.masked_values([-1, -1, 6, -1, 8], value=-1)
rng = aesara.shared(np.random.RandomState(2323), borrow=True)
with pytest.warns(ImputationWarning):
theta1 = Uniform("theta1", 0, 5, observed=obs1, rng=rng)
with pytest.warns(ImputationWarning):
theta2 = Normal("theta2", mu=theta1, observed=obs2, rng=rng)
assert "theta1_observed_interval__" in model.named_vars
assert "theta1_missing_interval__" in model.named_vars
assert isinstance(
model.rvs_to_values[model.named_vars["theta1_observed"]].tag.transform, Interval
)
prior_trace = sample_prior_predictive()
# Make sure the observed + missing combined deterministics have the
# same shape as the original observations vectors
assert prior_trace["theta1"].shape[-1] == obs1.shape[0]
assert prior_trace["theta2"].shape[-1] == obs2.shape[0]
# Make sure that the observed values are newly generated samples
assert np.all(np.var(prior_trace["theta1_observed"], 0) > 0.0)
assert np.all(np.var(prior_trace["theta2_observed"], 0) > 0.0)
# Make sure the missing parts of the combined deterministic matches the
# sampled missing and observed variable values
assert np.mean(prior_trace["theta1"][:, obs1.mask] - prior_trace["theta1_missing"]) == 0.0
assert np.mean(prior_trace["theta1"][:, ~obs1.mask] - prior_trace["theta1_observed"]) == 0.0
assert np.mean(prior_trace["theta2"][:, obs2.mask] - prior_trace["theta2_missing"]) == 0.0
assert np.mean(prior_trace["theta2"][:, ~obs2.mask] - prior_trace["theta2_observed"]) == 0.0
assert {"theta1", "theta2"} <= set(prior_trace.keys())
trace = sample(chains=1, draws=50, compute_convergence_checks=False)
assert np.all(0 < trace["theta1_missing"].mean(0))
assert np.all(0 < trace["theta2_missing"].mean(0))
assert "theta1" not in trace.varnames
assert "theta2" not in trace.varnames
# Make sure that the observed values are newly generated samples and that
# the observed and deterministic matche
pp_trace = sample_posterior_predictive(trace)
assert np.all(np.var(pp_trace["theta1"], 0) > 0.0)
assert np.all(np.var(pp_trace["theta2"], 0) > 0.0)
assert np.mean(pp_trace["theta1"][:, ~obs1.mask] - pp_trace["theta1_observed"]) == 0.0
assert np.mean(pp_trace["theta2"][:, ~obs2.mask] - pp_trace["theta2_observed"]) == 0.0
def test_step_categorical(self):
start, model, (mu, C) = simple_categorical()
unc = C ** .5
check = (('x', np.mean, mu, unc / 10.),
('x', np.std, unc, unc / 10.))
with model:
steps = (
CategoricalGibbsMetropolis(model.x, proposal='uniform'),
CategoricalGibbsMetropolis(model.x, proposal='proportional'),
)
for step in steps:
trace = sample(8000, step=step, start=start, model=model, random_seed=1)
yield self.check_stat, check, trace, step.__class__.__name__
def test_parallelized_chains_are_random(self):
with Model() as model:
x = Normal('x', 0, 1)
for stepper in TestPopulationSamplers.steppers:
step = stepper()
trace = sample(chains=4, draws=20, tune=0, step=DEMetropolis(),
parallelize=True)
samples = np.array(trace.get_values('x', combine=False))[:,5]
assert len(set(samples)) == 4, 'Parallelized {} ' \
'chains are identical.'.format(stepper)
pass
def test_step_elliptical_slice(self):
start, model, (K, L, mu, std, noise) = mv_prior_simple()
unc = noise ** 0.5
check = (('x', np.mean, mu, unc / 10.),
('x', np.std, std, unc / 10.))
with model:
steps = (
EllipticalSlice(prior_cov=K),
EllipticalSlice(prior_chol=L),
)
for step in steps:
trace = sample(5000, step=step, start=start, model=model, random_seed=1)
yield self.check_stat, check, trace, step.__class__.__name__
def test_step_categorical(self):
start, model, (mu, C) = simple_categorical()
unc = C ** 0.5
check = (("x", np.mean, mu, unc / 10.0), ("x", np.std, unc, unc / 10.0))
with model:
steps = (
CategoricalGibbsMetropolis(model.x, proposal="uniform"),
CategoricalGibbsMetropolis(model.x, proposal="proportional"),
)
for step in steps:
trace = sample(
8000, tune=0, step=step, start=start, model=model, random_seed=1
)
self.check_stat(check, trace, step.__class__.__name__)
def test_step_discrete(self):
if theano.config.floatX == "float32":
return # Cannot use @skip because it only skips one iteration of the yield
start, model, (mu, C) = mv_simple_discrete()
unc = np.diag(C) ** .5
check = (('x', np.mean, mu, unc / 10.),
('x', np.std, unc, unc / 10.))
with model:
steps = (
Metropolis(S=C, proposal_dist=MultivariateNormalProposal),
)
for step in steps:
trace = sample(20000, step=step, start=start, model=model, random_seed=1)
yield self.check_stat, check, trace, step.__class__.__name__
def test_step_discrete():
start, model, (mu, C) = mv_simple_discrete()
with model:
mh = Metropolis(S=C,
proposal_dist=MultivariateNormalProposal)
slicer = Slice()
steps = [mh]
unc = np.diag(C) ** .5
check = [('x', np.mean, mu, unc / 10.),
('x', np.std, unc, unc / 10.)]
for st in steps:
h = sample(20000, st, start, model=model, random_seed=1)
for (var, stat, val, bound) in check:
yield check_stat, repr(st), h, var, stat, val, bound
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("divergence after tuning" in warn for warn in warns)
or 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_step_continuous(self):
start, model, (mu, C) = mv_simple()
unc = np.diag(C) ** .5
check = (('x', np.mean, mu, unc / 10.),
('x', np.std, unc, unc / 10.))
with model:
steps = (
Slice(),
HamiltonianMC(scaling=C, is_cov=True, blocked=False),
NUTS(scaling=C, is_cov=True, blocked=False),
Metropolis(S=C, proposal_dist=MultivariateNormalProposal, blocked=True),
Slice(blocked=True),
HamiltonianMC(scaling=C, is_cov=True),
NUTS(scaling=C, is_cov=True),
CompoundStep([
HamiltonianMC(scaling=C, is_cov=True),
HamiltonianMC(scaling=C, is_cov=True, blocked=False)]),
)
for step in steps:
trace = sample(8000, step=step, start=start, model=model, random_seed=1)
yield self.check_stat, check, trace
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 in (NUTS, Slice, Metropolis):
trace = sample(100000, step=step_method(), progressbar=False)
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_bad_init(self):
with Model():
HalfNormal('a', sd=1, testval=-1, transform=None)
with pytest.raises(ValueError) as error:
sample(init=None)
error.match('Bad initial')
def test_bad_init_parallel(self):
with Model():
HalfNormal("a", sigma=1, testval=-1, transform=None)
with pytest.raises(ParallelSamplingError) as error:
sample(init=None, cores=2, random_seed=1)
error.match("Bad initial")