def test_kcenters_6():
# test with a custom metric when the input data isn't a list of numpy arrays
x = md.Trajectory(xyz=np.random.randn(100,1,3), topology=None)
# just get the sqeuclidean for the first atom along the first coordinate
metric = lambda target, ref, i: (target.xyz[:, 0, 0] - ref.xyz[i, 0, 0])**2
model1 = KCenters(n_clusters=10, metric=metric, random_state=0)
model1.fit([x])
model2 = KCenters(n_clusters=10, metric='sqeuclidean', random_state=0)
model2.fit([x.xyz[:, :, 0]])
eq(reduce(operator.add, model1.cluster_centers_).xyz[:, 0, 0],
model2.cluster_centers_[:, 0])
def test_kcenters_2():
# some data at (0,0), some data at (1,1) and some data at (0.5, 0.5)
data = [np.zeros((10,2)), np.ones((10,2)), 0.5*np.ones((10,2))]
m = KCenters(n_clusters=2, random_state=0)
m.fit(data)
# the centers should be [0,0], [1,1] (in either order). This
# assumes that the random state seeded the initial center at
# either (0,0) or (1,1). A different random state could have
# seeded the first cluster at [0.5, 0.5]
assert np.all(m.cluster_centers_ == np.array([[0,0], [1,1]])) or \
np.all(m.cluster_centers_ == np.array([[1,1], [0,0]]))
# the distances should be 0 or sqrt(2)/2
eq(np.unique(np.concatenate(m.distances_)), np.array([0, np.sqrt(2)/2]))
def test_kcenters_1():
# make sure all the shapes are correct of the fit parameters
m = KCenters(n_clusters=3)
m.fit([np.random.randn(23,2), np.random.randn(10,2)])
assert isinstance(m.labels_, list)
assert isinstance(m.distances_, list)
assert len(m.labels_) == 2
eq(m.cluster_centers_.shape, (3,2))
eq(m.labels_[0].shape, (23,))
eq(m.labels_[1].shape, (10,))
eq(m.distances_[0].shape, (23,))
eq(m.distances_[1].shape, (10,))
eq(m.fit_predict([np.random.randn(10, 2)])[0].shape, (10,))
