def _get_response_vector(source_index, num_sources, epsilon=0.):
response_vector = labels_to_one_hot(
np.array(source_index),
num_sources,
dtype=np.float64
)
response_vector = np.clip(response_vector, epsilon, 1.)
return response_vector
def test_cacgmm_sad_init(self):
samples = 10000
weight = np.array([0.3, 0.7])
num_classes, = weight.shape
covariance = np.array([
[[10, 1 + 1j, 1 + 1j], [1 - 1j, 5, 1], [1 - 1j, 1, 2]],
[[2, 0, 0], [0, 3, 0], [0, 0, 2]],
])
covariance /= np.trace(covariance, axis1=-2, axis2=-1)[..., None, None]
x, labels = sample_cacgmm(samples,
weight,
covariance,
return_label=True)
affiliations = labels_to_one_hot(labels, num_classes, axis=-2)
# test initialization
model = CACGMMTrainer().fit(
x,
initialization=affiliations,
covariance_norm='trace',
)
# test initialization with independent
model = CACGMMTrainer().fit(
np.array([x]),
initialization=np.array([affiliations]),
covariance_norm='trace',
)
# test initialization with independent and broadcasted initialization
model = CACGMMTrainer().fit(
np.array([x, x, x]),
initialization=np.array([affiliations]),
covariance_norm='trace',
)
# test initialization with independent
model = CACGMMTrainer().fit(
np.array([x, x]),
initialization=np.array([affiliations, affiliations]),
covariance_norm='trace',
)
def predict(self, x):
"""
Args:
x: Shape (N, D)
Returns: Affiliation with shape (K, N)
"""
N, D = x.shape
assert np.isrealobj(x), x.dtype
labels = self.kmeans.predict(x)
affiliations = labels_to_one_hot(labels,
self.kmeans.n_clusters,
axis=-2,
keepdims=False,
dtype=x.dtype)
assert affiliations.shape == (self.kmeans.n_clusters, N)
return affiliations