def test_augmented_L_construction(self):
# 5 LFs: a triangle, a connected edge to it, and a singleton source
n = 3
m = 5
k = 2
E = [(0,1), (1,2), (2,0), (0,3)]
L = np.array([
[1, 1, 1, 2, 1],
[1, 2, 2, 1, 0],
[1, 1, 1, 1, 0]
])
lm = LabelModel(m, k=k, deps=E)
L_aug = lm._get_augmented_label_matrix(L, offset=1, higher_order=True)
# Should have 22 columns:
# - 5 * 2 = 10 for the sources
# - 8 + 4 for the 3- and 2-clique resp. --> = 22
self.assertEqual(L_aug.shape, (3,22))
# Same as above but minus 2 abstains = 19 total nonzero entries
self.assertEqual(L_aug.sum(), 19)
# Next, check the singleton entries
for i in range(n):
for j in range(m):
if L[i,j] > 0:
self.assertEqual(L_aug[i, j * k + L[i,j] - 1], 1)
# Finally, check the clique entries
# Triangle clique
self.assertEqual(len(lm.c_tree.node[1]['members']), 3)
j = lm.c_tree.node[1]['start_index']
self.assertEqual(L_aug[0, j], 1)
self.assertEqual(L_aug[1, j + 3], 1)
self.assertEqual(L_aug[2, j], 1)
# Binary clique
self.assertEqual(len(lm.c_tree.node[2]['members']), 2)
j = lm.c_tree.node[2]['start_index']
self.assertEqual(L_aug[0, j+1], 1)
self.assertEqual(L_aug[1, j], 1)
self.assertEqual(L_aug[2, j], 1)
def _test_label_model(self, data, test_acc=True, mts=False):
if mts:
label_model = LabelModel(data.m, task_graph=data.task_graph,
p=data.p, deps=data.E)
else:
label_model = LabelModel(data.m, k=data.k, p=data.p, deps=data.E)
label_model.train(data.L, n_epochs=1000, print_every=200)
# Test parameter estimation error
c_probs_est = label_model.get_conditional_probs()
err = np.mean(np.abs(data.c_probs - c_probs_est))
print(f"Parameter Estimation Error={err}")
self.assertLess(err, 0.015)
# Test label prediction accuracy
if test_acc:
Y_pred = label_model.get_label_probs(data.L).argmax(axis=1) + 1
acc = np.where(data.Y == Y_pred, 1, 0).sum() / data.n
print(f"Label Prediction Accuracy={acc}")
self.assertGreater(acc, 0.95)
def _test_label_model(self, data, test_acc=True):
label_model = LabelModel(k=data.k, verbose=False)
label_model.train(
data.L,
deps=data.E,
class_balance=data.p,
n_epochs=1000,
print_every=200,
)
# Test parameter estimation error
c_probs_est = label_model.get_conditional_probs()
err = np.mean(np.abs(data.c_probs - c_probs_est))
self.assertLess(err, 0.025)
# Test label prediction accuracy
if test_acc:
Y_pred = label_model.predict_proba(data.L).argmax(axis=1) + 1
acc = np.where(data.Y == Y_pred, 1, 0).sum() / data.n
self.assertGreater(acc, 0.95)
def _test_label_model(self, data, test_acc=True):
label_model = LabelModel(k=data.k, verbose=False)
label_model.train_model(
data.L,
deps=data.E,
class_balance=data.p,
n_epochs=1000,
log_train_every=200,
)
# Test parameter estimation error
c_probs_est = label_model.get_conditional_probs()
err = np.mean(np.abs(data.c_probs - c_probs_est))
self.assertLess(err, 0.025)
# Test label prediction accuracy
if test_acc:
score = label_model.score((data.L, data.Y), verbose=False)
self.assertGreater(score, 0.95)
# Test against baseline
mv = MajorityLabelVoter()
mv_score = mv.score((data.L, data.Y), verbose=False)
self.assertGreater(score, mv_score)