def test_empirical_probabilities():
# In this test case we sample 10000 from randomly generated model and use
# samples to calculate empirical probabilities for every configuration.
# Then we explicitly evaluate theoretical probabilities of every state and
# assert that they are close to empirical probabilities.
model = tree_potts_model(gr_size=5, al_size=2)
num_samples = 10000
samples = model.sample(num_samples=num_samples, algorithm='tree_dp')
assert samples.shape == (num_samples, model.gr_size)
# Calculate empirical probabilities of states by counting how many
# times each state appeared in samples.
emp_proba = np.zeros(model.al_size**model.gr_size)
for sample in samples:
state_id = model.encode_state(sample)
emp_proba[state_id] += 1
emp_proba /= num_samples
# Calculate true probabilities of states by definition.
true_proba = [
model.evaluate(np.array(model.decode_state(i)))
for i in range(model.al_size**model.gr_size)
]
true_proba /= np.sum(true_proba)
assert np.max(true_proba - emp_proba) < 0.02
def test_tree():
model = tree_potts_model(gr_size=50, al_size=3, seed=0)
gt = model.infer(algorithm='tree_dp')
result = model.infer(algorithm='mean_field')
assert_results_close(result, gt, log_pf_tol=10.0, mp_mse_tol=0.03)
def test_infer_compare_with_pairwise_tree():
pw_model = tree_potts_model(gr_size=50, al_size=5, seed=0)
true_pf = np.exp(pw_model.infer(algorithm='tree_dp').log_pf)
nfg_model = inferlo.NormalFactorGraphModel.from_model(pw_model)
pf = infer_edge_elimination(nfg_model)
assert np.allclose(true_pf, pf)
def test_max_likelihood_potts_tree_1000x5():
libdai = LibDaiInterop()
if not libdai.is_libdai_ready():
return
model = tree_potts_model(1000, 5)
true_ml = model.max_likelihood()
libdai_ml = libdai.max_likelihood(model, "BP")
assert np.allclose(true_ml, libdai_ml)
def test_marg_probs_potts_tree_1000x5():
libdai = LibDaiInterop()
if not libdai.is_libdai_ready():
return
model = tree_potts_model(1000, 5)
true_result = model.infer()
libdai_result = libdai.infer(model, "BP")
assert_results_close(true_result, libdai_result, log_pf_tol=1e-8)
def test_big_tree():
gr_size = 1000
al_size = 5
j = np.ones((al_size, al_size)) + np.eye(al_size)
model = tree_potts_model(gr_size=gr_size,
al_size=al_size,
seed=111,
same_j=j,
zero_field=True)
result = model.infer(algorithm='tree_dp')
assert np.allclose(result.marg_prob, np.ones((gr_size, al_size)) / al_size)
def test_one_very_likely_state():
# In this test case we generate model on tree and then set very large
# fields such that they force values in all nodes to be equal to given
# value with probability very close to 1. Then we assert that this
# configuration was sampled at least 99% of times.
gr_size = 100
al_size = 5
model = tree_potts_model(gr_size=gr_size, al_size=al_size)
expected_configuration = np.random.choice(al_size, size=gr_size)
field = np.zeros((gr_size, al_size))
for i in range(gr_size):
field[i][expected_configuration[i]] = 100
model.set_field(field)
samples = model.sample(num_samples=100, algorithm='tree_dp')
good_count = sum(
[np.all(sample == expected_configuration) for sample in samples])
assert good_count >= 99
def test_tree_exact():
model = tree_potts_model(gr_size=50, al_size=3, seed=0)
assert_results_close(model.infer(algorithm='message_passing'),
model.infer(algorithm='tree_dp'))
def test_max_likelihood_tree_100x5():
model = tree_potts_model(gr_size=100, al_size=5, seed=0)
true_ml = model.max_likelihood(algorithm='tree_dp')
ml = max_likelihood_junction_tree(model)
assert np.allclose(ml, true_ml)
def test_inference_tree_100x5():
model = tree_potts_model(gr_size=100, al_size=5, seed=0)
ground_truth = model.infer(algorithm='tree_dp')
result = infer_junction_tree(model)
assert_results_close(result, ground_truth)
def test_tree_50x2():
model = tree_potts_model(gr_size=50, al_size=2, seed=0)
max_lh_gt = model.max_likelihood(algorithm='tree_dp')
max_lh = max_lh_path_dp(model)
assert np.allclose(max_lh, max_lh_gt)
def test_vert15_alph2():
model = tree_potts_model(gr_size=10, al_size=2, seed=123)
ground_truth = model.infer(algorithm='bruteforce')
result = model.infer(algorithm='tree_dp')
assert_results_close(result, ground_truth)
def test_tree_12x2():
for seed in range(5):
model = tree_potts_model(gr_size=12, al_size=2, seed=seed)
truth = model.max_likelihood(algorithm='bruteforce')
result = model.max_likelihood(algorithm='tree_dp')
assert np.allclose(truth, result)
def test_tree_5x3():
model = tree_potts_model(gr_size=5, al_size=3, seed=0)
truth = model.max_likelihood(algorithm='bruteforce')
result = model.max_likelihood(algorithm='tree_dp')
assert np.allclose(truth, result)
def test_tree():
model = tree_potts_model(gr_size=100, al_size=3)
true_result = model.infer(algorithm='tree_dp')
result_bp = BP.infer(model)
assert_results_close(true_result, result_bp, log_pf_tol=2e-9)
def test_tree_100x2():
model = tree_potts_model(gr_size=100, al_size=2, seed=0)
true_marg_probs = model.infer(algorithm='tree_dp').marg_prob
samples = model.sample(num_samples=10000, algorithm='tree_dp')
check_samples(samples=samples, true_marg_probs=true_marg_probs, tol=1e-4)