def test_fast_random_choice_output():
"""Test that the fast random choice outputs are accurate."""
n_sample = 10000
# few samples
weights = np.array([0.1, 0.9])
ret = np.array([fast_random_choice(weights) for _ in range(n_sample)])
assert 0.08 < sum(ret == 0) / n_sample < 0.12
assert 0.88 < sum(ret == 1) / n_sample < 0.92
# many samples
weights = np.array([*[0.02] * 20, *[0.3] * 2])
ret = np.array([fast_random_choice(weights) for _ in range(n_sample)])
assert 0.005 < sum(ret == 0) / n_sample < 0.04
assert 0.28 < sum(ret == 20) / n_sample < 0.32
def generate_valid_proposal(
t: int, m: np.ndarray, p: np.ndarray,
model_prior: RV, parameter_priors: List[Distribution],
model_perturbation_kernel: ModelPerturbationKernel,
transitions: List[Transition]):
"""Sample a parameter for a model.
Parameters
----------
t: Population index to generate for.
m: Indices of alive models.
p: Probabilities of alive models.
model_prior: The model prior.
parameter_priors: The parameter priors.
model_perturbation_kernel: The model perturbation kernel.
transitions: The transitions, one per model.
Returns
-------
(m_ss, theta_ss): Model, parameter.
"""
# first generation
if t == 0:
# sample from prior
m_ss = int(model_prior.rvs())
theta_ss = parameter_priors[m_ss].rvs()
return m_ss, theta_ss
# later generation
# counter
n_sample, n_sample_soft_limit = 0, 1000
# sample until the prior density is positive
while True:
if len(m) > 1:
index = fast_random_choice(p)
m_s = m[index]
m_ss = model_perturbation_kernel.rvs(m_s)
# theta_s is None if the population m_ss has died out.
# This can happen since the model_perturbation_kernel
# can return a model nr which has died out.
if m_ss not in m:
continue
else:
# only one model
m_ss = m[0]
theta_ss = Parameter(**transitions[m_ss].rvs().to_dict())
# check if positive under prior
if (model_prior.pmf(m_ss)
* parameter_priors[m_ss].pdf(theta_ss) > 0):
return m_ss, theta_ss
# unhealthy sampling detection
n_sample += 1
if n_sample == n_sample_soft_limit:
logger.warning(
"Unusually many (model, parameter) samples have prior "
"density zero. The transition might be inappropriate.")
def test_fast_random_choice():
"""Check that `fast_random_choice` delivers the promised benefit."""
nrep = 100000
ws = np.random.uniform(size=(nrep, 10))
ws /= ws.sum(axis=1, keepdims=True)
np.random.seed(0)
start = time()
for w in ws:
np.random.choice(len(w), p=w)
time_numpy = time() - start
print(f"Time numpy: {time_numpy}")
np.random.seed(0)
start = time()
for w in ws:
fast_random_choice(w)
time_fast = time() - start
print(f"Time fast_random_choice: {time_fast}")
assert time_fast < time_numpy
def test_fast_random_choice_basic():
"""Test the fast random choice function for various inputs."""
# run with many values (method 1)
weights = np.random.uniform(0, 1, size=10000)
weights /= weights.sum()
fast_random_choice(weights)
# run with few values (method 2)
weights = np.random.uniform(0, 1, size=5)
weights /= weights.sum()
fast_random_choice(weights)
# run with a single value
fast_random_choice(np.array([1]))
def test_fast_random_choice_errors():
"""Test the fast random choice function for invalid inputs."""
with pytest.raises(ValueError):
fast_random_choice(np.array([]))
# non-normalized weights for many values
weights = np.random.uniform(0, 1, size=1000)
weights /= weights.sum() / 2
with pytest.raises(ValueError):
fast_random_choice(weights)
# non-normalized weights for few values
with pytest.raises(ValueError):
# does not always raise, but will sometimes
for _ in range(100):
weights = np.random.uniform(0, 1, size=5)
weights /= 5
fast_random_choice(weights)
