def setUp(self):
self.points = np.array([[2.5, 0.1], [2, 0.2], [3, 0.1], [4, 0.2],
[0.1, 2.5], [0.2, 2], [0.1, 3], [0.2, 4]],
dtype=np.float32)
self.num_points = self.points.shape[0]
self.true_centers = np.array([
normalize(
np.mean(normalize(self.points)[0:4, :], axis=0,
keepdims=True))[0],
normalize(
np.mean(normalize(self.points)[4:, :], axis=0,
keepdims=True))[0]
],
dtype=np.float32)
self.true_assignments = np.array([0] * 4 + [1] * 4)
self.true_score = len(self.points) - np.tensordot(
normalize(self.points), self.true_centers[self.true_assignments])
self.num_centers = 2
self.kmeans = kmeans_lib.KMeansClustering(
self.num_centers,
initial_clusters=kmeans_lib.KMeansClustering.RANDOM_INIT,
distance_metric=kmeans_lib.KMeansClustering.COSINE_DISTANCE,
use_mini_batch=self.use_mini_batch,
mini_batch_steps_per_iteration=self.mini_batch_steps_per_iteration,
config=self.config(3))
def _kmeans(self, relative_tolerance=None):
return kmeans_lib.KMeansClustering(
self.num_centers,
initial_clusters=self.initial_clusters,
distance_metric=kmeans_lib.KMeansClustering.
SQUARED_EUCLIDEAN_DISTANCE,
use_mini_batch=self.use_mini_batch,
mini_batch_steps_per_iteration=self.mini_batch_steps_per_iteration,
random_seed=24,
relative_tolerance=relative_tolerance)
def test_predict_kmeans_plus_plus(self):
# Most points are concentrated near one center. KMeans++ is likely to find
# the less populated centers.
points = np.array([[2.5, 3.5], [2.5, 3.5], [-2, 3], [-2, 3], [-3, -3],
[-3.1, -3.2], [-2.8, -3.], [-2.9, -3.1],
[-3., -3.1], [-3., -3.1], [-3.2, -3.], [-3., -3.]],
dtype=np.float32)
true_centers = np.array([
normalize(np.mean(normalize(points)[0:2, :], axis=0,
keepdims=True))[0],
normalize(np.mean(normalize(points)[2:4, :], axis=0,
keepdims=True))[0],
normalize(np.mean(normalize(points)[4:, :], axis=0,
keepdims=True))[0]
],
dtype=np.float32)
true_assignments = [0] * 2 + [1] * 2 + [2] * 8
true_score = len(points) - np.tensordot(normalize(points),
true_centers[true_assignments])
kmeans = kmeans_lib.KMeansClustering(
3,
initial_clusters=self.initial_clusters,
distance_metric=kmeans_lib.KMeansClustering.COSINE_DISTANCE,
use_mini_batch=self.use_mini_batch,
mini_batch_steps_per_iteration=self.mini_batch_steps_per_iteration,
config=self.config(3))
kmeans.train(input_fn=lambda: (constant_op.constant(points), None),
steps=30)
centers = normalize(kmeans.cluster_centers())
self.assertAllClose(sorted(centers.tolist()),
sorted(true_centers.tolist()),
atol=1e-2)
def _input_fn():
return (input_lib.limit_epochs(constant_op.constant(points),
num_epochs=1), None)
assignments = list(kmeans.predict_cluster_index(input_fn=_input_fn))
self.assertAllClose(centers[assignments],
true_centers[true_assignments],
atol=1e-2)
score = kmeans.score(
input_fn=lambda: (constant_op.constant(points), None))
self.assertAllClose(score, true_score, atol=1e-2)
def test_kmeans_plus_plus_batch_too_small(self):
points = np.array([[1, 2], [3, 4], [5, 6], [7, 8], [9, 0]],
dtype=np.float32)
kmeans = kmeans_lib.KMeansClustering(
num_clusters=points.shape[0],
initial_clusters=kmeans_lib.KMeansClustering.KMEANS_PLUS_PLUS_INIT,
distance_metric=kmeans_lib.KMeansClustering.
SQUARED_EUCLIDEAN_DISTANCE,
use_mini_batch=True,
mini_batch_steps_per_iteration=100,
random_seed=24,
relative_tolerance=None)
with self.assertRaisesOpError(AssertionError):
kmeans.train(input_fn=self.input_fn(batch_size=4,
points=points,
randomize=False),
steps=1)
def test_kmeans_plus_plus_batch_just_right(self):
points = np.array([[1, 2]], dtype=np.float32)
kmeans = kmeans_lib.KMeansClustering(
num_clusters=points.shape[0],
initial_clusters=kmeans_lib.KMeansClustering.KMEANS_PLUS_PLUS_INIT,
distance_metric=kmeans_lib.KMeansClustering.
SQUARED_EUCLIDEAN_DISTANCE,
use_mini_batch=True,
mini_batch_steps_per_iteration=100,
random_seed=24,
relative_tolerance=None)
kmeans.train(input_fn=self.input_fn(batch_size=1,
points=points,
randomize=False),
steps=1)
clusters = kmeans.cluster_centers()
self.assertAllEqual(points, clusters)
def _fit(self, num_iters=10):
scores = []
start = time.time()
for i in range(num_iters):
print('Starting tensorflow KMeans: %d' % i)
tf_kmeans = kmeans_lib.KMeansClustering(
self.num_clusters,
initial_clusters=kmeans_lib.KMeansClustering.
KMEANS_PLUS_PLUS_INIT,
kmeans_plus_plus_num_retries=int(
math.log(self.num_clusters) + 2),
random_seed=i * 42,
relative_tolerance=1e-6,
config=self.config(3))
tf_kmeans.train(input_fn=lambda:
(constant_op.constant(self.points), None),
steps=50)
_ = tf_kmeans.cluster_centers()
scores.append(
tf_kmeans.score(input_fn=lambda:
(constant_op.constant(self.points), None)))
self._report(num_iters, start, time.time(), scores)
def test_monitor(self):
if self.use_mini_batch:
# We don't test for use_mini_batch case since the loss value can be noisy.
return
kmeans = kmeans_lib.KMeansClustering(
self.num_centers,
initial_clusters=self.initial_clusters,
distance_metric=kmeans_lib.KMeansClustering.
SQUARED_EUCLIDEAN_DISTANCE,
use_mini_batch=self.use_mini_batch,
mini_batch_steps_per_iteration=self.mini_batch_steps_per_iteration,
config=self.config(14),
random_seed=12,
relative_tolerance=1e-4)
kmeans.train(
input_fn=self.input_fn(),
# Force it to train until the relative tolerance monitor stops it.
steps=None)
score = kmeans.score(input_fn=self.input_fn(
batch_size=self.num_points))
self.assertNear(self.true_score, score, self.true_score * 0.01)
def test_queues(self):
kmeans = kmeans_lib.KMeansClustering(5)
kmeans.train(input_fn=self.input_fn(), steps=1)