def execute(cls, ctx, op: "KMeansPlusPlusInit"):
try:
from sklearn.cluster._kmeans import _k_init
except ImportError: # pragma: no cover
from sklearn.cluster.k_means_ import _k_init
(x, x_squared_norms), device_id, _ = as_same_device(
[ctx[inp.key] for inp in op.inputs],
device=op.device,
ret_extra=True)
with device(device_id):
ctx[op.outputs[0].key] = _k_init(x, op.n_clusters, x_squared_norms,
op.state, op.n_local_trials)
def fit(self, X):
self.high_dim_data = X
self.highdim = self.high_dim_data.shape[1]
print("Projecting {}-dimensional data to {} dimensions ...".format(
self.highdim, self.n_dims))
starttime = time.time()
self.proj_vectors = [
rand_vector(self.highdim) for n in range(self.n_dims)
]
if self.n_dims == self.highdim:
self.low_dim_data = np.copy(X)
else:
self.low_dim_data = [
np.array(self.project(x)) for x in self.high_dim_data
]
print("Projection completed in {} seconds.".format(time.time() -
starttime))
self.datamax = [max(i) for i in zip(*self.low_dim_data)]
self.datamin = [min(i) for i in zip(*self.low_dim_data)]
x_squared_norms = row_norms(self.low_dim_data, squared=True)
starttime = time.time()
cents = _k_init(np.array(self.low_dim_data), self.n_clusters,
x_squared_norms, RandomState())
kmcents = _k_init(np.array(self.high_dim_data), self.n_clusters,
row_norms(self.high_dim_data, squared=True),
RandomState())
print("Initialization of centers completed in {} seconds.".format(
time.time() - starttime))
self.JLKcenters = [HDNode(center=cent) for cent in cents]
self.KMcenters = [HDNode(center=np.copy(cent)) for cent in kmcents]
def _get_initialization(self):
inits = list()
def squared_euclidean(x):
return np.inner(x, x)
squared_norms = np.apply_along_axis(func1d=squared_euclidean,
axis=1,
arr=self.X_norm)
for seed in self.seeds:
init = _k_init(X=self.X_norm,
n_clusters=self.n_clusters,
x_squared_norms=squared_norms,
random_state=seed)
inits.append(init)
self.inits = inits
def init_model(
lm_model,
tokenizer,
data_loader,
n_clusters,
embedding_extractor=concat_cls_n_hidden_states,
device='cpu',
random_state=np.random.RandomState(42),
**kwargs,
):
initial_embeddings = []
labels = []
for batch_texts, batch_labels in data_loader:
inputs = tokenizer(list(batch_texts),
return_tensors='pt',
padding=True,
truncation=True)
inputs = inputs.to(device)
with torch.no_grad():
outputs = lm_model.base_model(**inputs)
extracted_embeddings = embedding_extractor(
outputs).cpu().detach().numpy()
initial_embeddings.append(extracted_embeddings)
labels.extend(batch_labels.numpy().astype('int'))
initial_embeddings = np.vstack(initial_embeddings)
initial_centroids = _k_init(initial_embeddings,
n_clusters=n_clusters,
x_squared_norms=row_norms(initial_embeddings,
squared=True),
random_state=random_state)
model = ClusterLM(lm_model=lm_model,
tokenizer=tokenizer,
embedding_extractor=embedding_extractor,
initial_centroids=torch.from_numpy(initial_centroids),
device=device,
**kwargs)
return model, initial_centroids, initial_embeddings
def get_initialization_centroids(self, input_data):
"""
:param input_data: Matrix shape nxd (n: number of observations; d: dimensionality of observations)
:return:
"""
if self["--initialization"] == "random":
return np.random.normal(
(self["--nb-cluster"], input_data.shape[1]))
elif self["--initialization"] == "uniform_sampling":
return input_data[np.random.permutation(
input_data.shape[0])[:self["--nb-cluster"]]]
elif self["--initialization"] == "kmeans++":
seed = np.random.RandomState(self["--seed"])
x_squared_norms = row_norms(input_data, squared=True)
centers = _k_init(input_data,
self["--nb-cluster"],
x_squared_norms,
random_state=seed)
return centers
else:
raise NotImplementedError("Unknown initialization.")
def _kmeans_plusplus(*args, **kwargs):
return _k_init(*args, **kwargs), None
from sklearn.cluster import KMeans import matplotlib.pyplot as plt import numpy as np from sklearn.cluster._kmeans import _k_init from sklearn.utils.extmath import row_norms from sklearn.datasets import make_blobs k = 4 seed = np.random.RandomState(0) X, y = make_blobs(1000, centers=k) x_squared_norms = row_norms(X, squared=True) centers = _k_init(X, k, x_squared_norms, random_state=seed) random_centers = X[seed.permutation(X.shape[0])[:k]] plt.scatter(X[:, 0], X[:, 1], c="c") plt.scatter(centers[:, 0], centers[:, 1], c="g", label="kmeans++") plt.scatter(random_centers[:, 0], random_centers[:, 1], c="r", label="random uniform") plt.legend() plt.show()
