def _init_w(self, V, X):
"""
Initialize the topics W.
If self.init='k-means++', we use the init method of
sklearn.cluster.KMeans.
If self.init='random', topics are initialized with a Gamma
distribution.
If self.init='k-means', topics are initialized with a KMeans on the
n-grams counts.
"""
if self.init == 'k-means++':
if LooseVersion(sklearn_version) < LooseVersion('0.24'):
W = _k_init(
V, self.n_components,
x_squared_norms=row_norms(V, squared=True),
random_state=self.random_state,
n_local_trials=None) + .1
else:
W, _ = kmeans_plusplus(
V, self.n_components,
x_squared_norms=row_norms(V, squared=True),
random_state=self.random_state,
n_local_trials=None)
W = W + .1 # To avoid restricting topics to few n-grams only
elif self.init == 'random':
W = self.random_state.gamma(
shape=self.gamma_shape_prior, scale=self.gamma_scale_prior,
size=(self.n_components, self.n_vocab))
elif self.init == 'k-means':
prototypes = get_kmeans_prototypes(
X, self.n_components, random_state=self.random_state)
W = self.ngrams_count_.transform(prototypes).A + .1
if self.add_words:
W2 = self.word_count_.transform(prototypes).A + .1
W = np.hstack((W, W2))
# if k-means doesn't find the exact number of prototypes
if W.shape[0] < self.n_components:
if LooseVersion(sklearn_version) < LooseVersion('0.24'):
W2 = _k_init(
V, self.n_components - W.shape[0],
x_squared_norms=row_norms(V, squared=True),
random_state=self.random_state,
n_local_trials=None) + .1
else:
W2, _ = kmeans_plusplus(
V, self.n_components - W.shape[0],
x_squared_norms=row_norms(V, squared=True),
random_state=self.random_state,
n_local_trials=None)
W2 = W2 + .1
W = np.concatenate((W, W2), axis=0)
else:
raise AttributeError(
'Initialization method %s does not exist.' % self.init)
W /= W.sum(axis=1, keepdims=True)
A = np.ones((self.n_components, self.n_vocab)) * 1e-10
B = A.copy()
return W, A, B
def test_kmeans_plusplus_dataorder():
# Check that memory layout does not effect result
centers_c, _ = kmeans_plusplus(X, n_clusters, random_state=0)
X_fortran = np.asfortranarray(X)
centers_fortran, _ = kmeans_plusplus(X_fortran, n_clusters, random_state=0)
assert_allclose(centers_c, centers_fortran)
def test_kmeans_plusplus_norms(x_squared_norms):
# Check that defining x_squared_norms returns the same as default=None.
centers, indices = kmeans_plusplus(X,
n_clusters,
x_squared_norms=x_squared_norms)
assert_allclose(X[indices], centers)
def __kmeans_plus_plus_init(self, X):
# Initial centers
# It chooses the first centroid randomly and the next ones
# using a weighted probability p i = cost i /SUM( cost i ), where cost i is
# the squared distance of the data point x i to its nearest centroids.
self.cluster_centers_, _ = kmeans_plusplus(X,
n_clusters=self.n_clusters)
def kmeans_pp(inst):
random_state = check_random_state(None)
x_squared_norms = row_norms(inst.X, squared=True)
centers, indices = kmeans_plusplus(inst.X,
inst.p,
random_state=random_state,
x_squared_norms=x_squared_norms)
return indices
def initialize_arg(self, X):
'''
Initialize EM algorithm
'''
n_samples, n_features = X.shape
log_pi = np.log(np.full((1, self.n_clusters), 1 / self.n_clusters))
low, _ = kmeans_plusplus(X, self.n_clusters)
high = low
scale = np.ones((self.n_clusters, n_features)) / self.n_clusters
return {"log_pi": log_pi, "low": low, "high": high, "scale": scale}
def test_kmeans_plusplus_output(data, dtype):
# Check for the correct number of seeds and all positive values
data = data.astype(dtype)
centers, indices = kmeans_plusplus(data, n_clusters)
# Check there are the correct number of indices and that all indices are
# positive and within the number of samples
assert indices.shape[0] == n_clusters
assert (indices >= 0).all()
assert (indices <= data.shape[0]).all()
# Check for the correct number of seeds and that they are bound by the data
assert centers.shape[0] == n_clusters
assert (centers.max(axis=0) <= data.max(axis=0)).all()
assert (centers.min(axis=0) >= data.min(axis=0)).all()
# Check that indices correspond to reported centers
# Use X for comparison rather than data, test still works against centers
# calculated with sparse data.
assert_allclose(X[indices].astype(dtype), centers)
def test_kmeans_plusplus_wrong_params(param, match):
with pytest.raises(ValueError, match=match):
kmeans_plusplus(X, n_clusters, **param)
from sklearn.cluster import kmeans_plusplus
from sklearn.datasets import make_blobs
import matplotlib.pyplot as plt
# Generate sample data
n_samples = 4000
n_components = 4
X, y_true = make_blobs(n_samples=n_samples,
centers=n_components,
cluster_std=0.60,
random_state=0)
X = X[:, ::-1]
# Calculate seeds from kmeans++
centers_init, indices = kmeans_plusplus(X, n_clusters=4,
random_state=0)
# Plot init seeds along side sample data
plt.figure(1)
colors = ['#4EACC5', '#FF9C34', '#4E9A06', 'm']
for k, col in enumerate(colors):
cluster_data = y_true == k
plt.scatter(X[cluster_data, 0], X[cluster_data, 1],
c=col, marker='.', s=10)
plt.scatter(centers_init[:, 0], centers_init[:, 1], c='b', s=50)
plt.title("K-Means++ Initialization")
plt.xticks([])
plt.yticks([])
#plt.show()
def update(self, g: RandomWalkGraph, lr=None) -> int:
"""
update the clusters on a single batch of random walks
:param g: the graph being clustered
:param lr: (optional) learning rate parameter. If specified, overrides the object's lr property
:return: number of float values stored in the weight matrix
"""
lr = lr or self.lr
eq_sample_count = 1 / lr - 1
# get walks
walks = [
set(
g.unweighted_random_walk(
length=random.randint(self.min_len, self.max_len)))
for _ in range(self.batch_size)
]
# assign new nodes
all_states = set().union(*[walk for walk in walks])
unmapped = all_states - set(self.map)
if len(unmapped) > 0:
self.map.update(
dict(
zip(unmapped,
range(len(self.map),
len(self.map) + len(unmapped)))))
if self.centers is not None:
self.centers.resize(self.n_clusters, len(self.map))
# convert walks to binary
x = self._walks_to_matrix(walks)
# initialize if this is the first iteration
if self.centers is None:
self.centers = sparse.csr_matrix(
normalize(kmeans_plusplus(x, self.n_clusters)[0],
norm='l2',
axis=1,
copy=False))
# compute distances (dot products)
dots = self.centers.dot(x.T)
winners = dots.argmax(axis=0).A1
# compute data means by cluster
onehot_labels = sparse.csr_matrix(
(np.ones(self.batch_size), (winners, np.arange(self.batch_size))),
shape=(self.n_clusters, self.batch_size))
cluster_means = onehot_labels.dot(x)
cluster_counts = onehot_labels.sum(axis=1).A1
row_indices, _ = cluster_means.nonzero()
cluster_means.data /= cluster_counts[row_indices]
# derive a learning rate for each cluster, based on number of samples
lr_by_cluster = cluster_counts / (eq_sample_count + cluster_counts)
# do weighted average
cluster_means.data *= lr_by_cluster[row_indices]
row_indices, _ = self.centers.nonzero()
self.centers.data *= (1 - lr_by_cluster)[row_indices]
self.centers += cluster_means
# prune
self.centers.data[self.centers.data < self.drop_threshold] = 0
self.centers.eliminate_zeros()
# reproject cluster centers
self.centers = normalize(self.centers, norm='l2', axis=1, copy=False)
return len(self.centers.data)
