def learn_mmc_metric(X_test):
mmc_dict = dict()
for respondent_id in range(1, 21):
y_test = pd.DataFrame(
pd.read_pickle(r'../data/HCON/HCON_long_lik.pkl')
[respondent_id]).values.reshape(-1, 1)
mask = (y_test[None] == y_test[:, None])[:, :, 0]
a, b = np.nonzero(np.triu(mask, k=1)) # similarity pairs
c, d = np.nonzero(np.triu(~mask, k=1)) # dissimilarity pairs
mmc = MMC(convergence_threshold=0.001)
try:
mmc.fit(X_test.values, (a, b, c, d))
L = mmc.transform(np.diag(np.ones(9)))
M = np.dot(L, L.T)
except ValueError:
# it should be converged anyway,
# if the ValueError happens, there is some bad patterns of the input
print(
'R%d has no non-trivial dissimilarity constraints given for MMC.'
% respondent_id)
M = 0.01 * np.diag(np.ones(9))
mmc_dict['R%d' % respondent_id] = M * 100
print('R:%2d' % respondent_id, ' First Row of MMC Mahalanobis Matrix:',
(M[0] * 100).round(3))
return mmc_dict
def fit(self, X, y=None, ml=[], cl=[]):
X_transformed = X
if ml and cl:
# ml_graph, cl_graph, _ = preprocess_constraints(ml, cl, X.shape[0])
#
# ml, cl = [], []
# for i, constraints in ml_graph.items():
# for j in constraints:
# ml.append((i, j))
#
# for i, constraints in cl_graph.items():
# for j in constraints:
# cl.append((i, j))
constraints = [np.array(lst) for lst in [*zip(*ml), *zip(*cl)]]
mmc = MMC(diagonal=self.diagonal)
mmc.fit(X, constraints=constraints)
X_transformed = mmc.transform(X)
kmeans = KMeans(n_clusters=self.n_clusters, init='random', max_iter=self.max_iter)
kmeans.fit(X_transformed)
self.labels_ = kmeans.labels_
return self
def runMMC():
# Run MMC
from metric_learn import MMC
"""
Learn MMC (Mahalanobis Metrics for Clustering) Model
"""
mmc = MMC()
mmc.fit(pairs, y) # learn the MMC model
print("Mahalanobis Matrix : ", mmc.get_mahalanobis_matrix())
def fit(self, X, y=None, constraints=None):
mmc = MMC(diagonal=self.diagonal)
mmc.fit(X, constraints=constraints)
X_transformed = mmc.transform(X)
kmeans = KMeans(n_clusters=self.n_clusters,
init='random',
max_iter=self.max_iter)
kmeans.fit(X_transformed)
self.labels_ = kmeans.labels_
return self
def main(args):
print("Deriving similar/dissimilar constraints for metric learning.")
with gzip.open(args.transfer_acc, "rb") as fr:
# transer_acc[tgt][src]: accuracy of src->tgt
transfer_acc = pickle.load(fr)
_mean = {
l: mean(list(transfer_acc[l].values()))
for l in transfer_acc.keys()
}
_std = {
l: stdev(list(transfer_acc[l].values()))
for l in transfer_acc.keys()
}
alpha = 0.5
sim_pairs = []
dissim_pairs = []
meta_langs = list(transfer_acc.keys())
for i in range(len(meta_langs)):
for j in range(i + 1, len(meta_langs)):
l1 = meta_langs[i]
l2 = meta_langs[j]
if transfer_acc[l1][l2] > _mean[l1] + alpha * _std[l1] and \
transfer_acc[l2][l1] > _mean[l2] + alpha * _std[l2]:
sim_pairs.append([l1, l2])
elif transfer_acc[l1][l2] < _mean[l1] - alpha * _std[l1] and \
transfer_acc[l2][l1] < _mean[l2] - alpha * _std[l2]:
dissim_pairs.append([l1, l2])
# constraints: [simA, simB, dissimA, dissimB]
constraints = list(zip(*sim_pairs)) + list(zip(*dissim_pairs))
constraints = [
list(map(lambda l: meta_langs.index(l), lst)) for lst in constraints
]
constraints = [np.array(x) for x in constraints]
print("Mahalanobis metric learning.")
with gzip.open(args.feature_path, "rb") as fr:
typology_vec = pickle.load(fr)
meta_X = np.array([typology_vec[l] for l in meta_langs])
mmc = MMC()
mmc.fit(meta_X, constraints)
print("Apply the learned metric to the full typology vector space.")
all_langs = list(typology_vec.keys())
X = np.array([typology_vec[l] for l in all_langs])
X = mmc.transform(X).tolist()
typology_vec_transformed = {
all_langs[i]: X[i]
for i in range(len(all_langs))
}
with gzip.open(args.output_file, "wb") as fw:
pickle.dump(typology_vec_transformed, fw)
def test_iris(self):
# Generate full set of constraints for comparison with reference implementation
n = self.iris_points.shape[0]
mask = (self.iris_labels[None] == self.iris_labels[:, None])
a, b = np.nonzero(np.triu(mask, k=1))
c, d = np.nonzero(np.triu(~mask, k=1))
# Full metric
mmc = MMC(convergence_threshold=0.01)
mmc.fit(self.iris_points, [a, b, c, d])
expected = [[0.000514, 0.000868, -0.001195, -0.001703],
[0.000868, 0.001468, -0.002021, -0.002879],
[-0.001195, -0.002021, 0.002782, 0.003964],
[-0.001703, -0.002879, 0.003964, 0.005648]]
assert_array_almost_equal(expected, mmc.metric(), decimal=6)
# Diagonal metric
mmc = MMC(diagonal=True)
mmc.fit(self.iris_points, [a, b, c, d])
expected = [0, 0, 1.210220, 1.228596]
assert_array_almost_equal(np.diag(expected), mmc.metric(), decimal=6)
# Supervised Full
mmc = MMC_Supervised()
mmc.fit(self.iris_points, self.iris_labels)
csep = class_separation(mmc.transform(), self.iris_labels)
self.assertLess(csep, 0.15)
# Supervised Diagonal
mmc = MMC_Supervised(diagonal=True)
mmc.fit(self.iris_points, self.iris_labels)
csep = class_separation(mmc.transform(), self.iris_labels)
self.assertLess(csep, 0.2)
def test_iris(self):
# Generate full set of constraints for comparison with reference implementation
mask = (self.iris_labels[None] == self.iris_labels[:, None])
a, b = np.nonzero(np.triu(mask, k=1))
c, d = np.nonzero(np.triu(~mask, k=1))
# Full metric
mmc = MMC(convergence_threshold=0.01)
mmc.fit(self.iris_points, [a, b, c, d])
expected = [[+0.00046504, +0.00083371, -0.00111959, -0.00165265],
[+0.00083371, +0.00149466, -0.00200719, -0.00296284],
[-0.00111959, -0.00200719, +0.00269546, +0.00397881],
[-0.00165265, -0.00296284, +0.00397881, +0.00587320]]
assert_array_almost_equal(expected, mmc.metric(), decimal=6)
# Diagonal metric
mmc = MMC(diagonal=True)
mmc.fit(self.iris_points, [a, b, c, d])
expected = [0, 0, 1.21045968, 1.22552608]
assert_array_almost_equal(np.diag(expected), mmc.metric(), decimal=6)
# Supervised Full
mmc = MMC_Supervised()
mmc.fit(self.iris_points, self.iris_labels)
csep = class_separation(mmc.transform(), self.iris_labels)
self.assertLess(csep, 0.15)
# Supervised Diagonal
mmc = MMC_Supervised(diagonal=True)
mmc.fit(self.iris_points, self.iris_labels)
csep = class_separation(mmc.transform(), self.iris_labels)
self.assertLess(csep, 0.2)
def test_iris(self):
# Generate full set of constraints for comparison with reference implementation
mask = (self.iris_labels[None] == self.iris_labels[:, None])
a, b = np.nonzero(np.triu(mask, k=1))
c, d = np.nonzero(np.triu(~mask, k=1))
# Full metric
mmc = MMC(convergence_threshold=0.01)
mmc.fit(self.iris_points, [a, b, c, d])
expected = [[+0.00046504, +0.00083371, -0.00111959, -0.00165265],
[+0.00083371, +0.00149466, -0.00200719, -0.00296284],
[-0.00111959, -0.00200719, +0.00269546, +0.00397881],
[-0.00165265, -0.00296284, +0.00397881, +0.00587320]]
assert_array_almost_equal(expected, mmc.metric(), decimal=6)
# Diagonal metric
mmc = MMC(diagonal=True)
mmc.fit(self.iris_points, [a, b, c, d])
expected = [0, 0, 1.21045968, 1.22552608]
assert_array_almost_equal(np.diag(expected), mmc.metric(), decimal=6)
# Supervised Full
mmc = MMC_Supervised()
mmc.fit(self.iris_points, self.iris_labels)
csep = class_separation(mmc.transform(), self.iris_labels)
self.assertLess(csep, 0.15)
# Supervised Diagonal
mmc = MMC_Supervised(diagonal=True)
mmc.fit(self.iris_points, self.iris_labels)
csep = class_separation(mmc.transform(), self.iris_labels)
self.assertLess(csep, 0.2)
def learn_distance_metric(distances, pairs_per_prototype=100,
test_size=0.5,
return_features=False,
return_pairs=False):
feature_pipeline = Pipeline([
('dates', DateFeatureTransformer()),
('features', MMCFeatureTransformer()),
])
features = feature_pipeline.fit_transform(distances)
pairs = create_mmc_pairs(distances, pairs_per_prototype=pairs_per_prototype)
X_train, X_test, y_train, y_test = train_test_split(pairs[:, :2], pairs[:, -1],
shuffle=True, stratify=pairs[:, -1], test_size=test_size
)
mmc = MMC(preprocessor=np.array(features, dtype=np.float))
mmc = mmc.fit(X_train, y_train)
score = f1_score(y_test, mmc.predict(X_test), average='weighted')
return SimpleNamespace(
score=score,
metric_components=mmc.components_.transpose(),
features=None if not return_features else features,
pairs=None if not return_pairs else pairs
)
def test_iris(self):
# Generate full set of constraints for comparison with reference
# implementation
mask = self.iris_labels[None] == self.iris_labels[:, None]
a, b = np.nonzero(np.triu(mask, k=1))
c, d = np.nonzero(np.triu(~mask, k=1))
# Full metric
n_features = self.iris_points.shape[1]
mmc = MMC(convergence_threshold=0.01, init=np.eye(n_features) / 10)
mmc.fit(*wrap_pairs(self.iris_points, [a, b, c, d]))
expected = [[+0.000514, +0.000868, -0.001195, -0.001703],
[+0.000868, +0.001468, -0.002021, -0.002879],
[-0.001195, -0.002021, +0.002782, +0.003964],
[-0.001703, -0.002879, +0.003964, +0.005648]]
assert_array_almost_equal(expected,
mmc.get_mahalanobis_matrix(),
decimal=6)
# Diagonal metric
mmc = MMC(diagonal=True)
mmc.fit(*wrap_pairs(self.iris_points, [a, b, c, d]))
expected = [0, 0, 1.210220, 1.228596]
assert_array_almost_equal(np.diag(expected),
mmc.get_mahalanobis_matrix(),
decimal=6)
# Supervised Full
mmc = MMC_Supervised()
mmc.fit(self.iris_points, self.iris_labels)
csep = class_separation(mmc.transform(self.iris_points),
self.iris_labels)
self.assertLess(csep, 0.15)
# Supervised Diagonal
mmc = MMC_Supervised(diagonal=True)
mmc.fit(self.iris_points, self.iris_labels)
csep = class_separation(mmc.transform(self.iris_points),
self.iris_labels)
self.assertLess(csep, 0.2)
def test_iris(self):
# Generate full set of constraints for comparison with reference implementation
n = self.iris_points.shape[0]
mask = (self.iris_labels[None] == self.iris_labels[:,None])
a, b = np.nonzero(np.triu(mask, k=1))
c, d = np.nonzero(np.triu(~mask, k=1))
# Full metric
mmc = MMC(convergence_threshold=0.01)
mmc.fit(*wrap_pairs(self.iris_points, [a,b,c,d]))
expected = [[+0.000514, +0.000868, -0.001195, -0.001703],
[+0.000868, +0.001468, -0.002021, -0.002879],
[-0.001195, -0.002021, +0.002782, +0.003964],
[-0.001703, -0.002879, +0.003964, +0.005648]]
assert_array_almost_equal(expected, mmc.get_mahalanobis_matrix(),
decimal=6)
# Diagonal metric
mmc = MMC(diagonal=True)
mmc.fit(*wrap_pairs(self.iris_points, [a,b,c,d]))
expected = [0, 0, 1.210220, 1.228596]
assert_array_almost_equal(np.diag(expected), mmc.get_mahalanobis_matrix(),
decimal=6)
# Supervised Full
mmc = MMC_Supervised()
mmc.fit(self.iris_points, self.iris_labels)
csep = class_separation(mmc.transform(self.iris_points), self.iris_labels)
self.assertLess(csep, 0.15)
# Supervised Diagonal
mmc = MMC_Supervised(diagonal=True)
mmc.fit(self.iris_points, self.iris_labels)
csep = class_separation(mmc.transform(self.iris_points), self.iris_labels)
self.assertLess(csep, 0.2)
"""
from metric_learn import MMC
pairs = [[[1.2, 7.5], [1.3, 1.5]], [[6.4, 2.6], [6.2, 9.7]],
[[1.3, 4.5], [3.2, 4.6]], [[6.2, 5.5], [5.4, 5.4]]]
# in this task we want points where the first feature is close to be closer to each other,
# no matter how close the second feature is
y = [1, 1, -1, -1]
"""
Learn MMC (Mahalanobis Metrics for Clustering) Model
"""
mmc = MMC()
mmc.fit(pairs, y) # learn the MMC model
"""
Return the decision function used to classify the pairs
"""
print("debug 1: ", mmc.decision_function(pairs))
"""
Returns a copy of the Mahalanobis matrix learned by the metric learner
"""
print("debug 2: ", mmc.get_mahalanobis_matrix())
"""
Returns a function that takes as input two 1D arrays and outputs the learned metric score on these two points.
"""
f = mmc.get_metric()
print("debug 3: ", f)
"""
Predicts the learned metric between input pairs