def affinity(simdata):
# this should handle a lot of different argument definitions
# just test two (single list, individual arguments) to compare outputs
aff = compute.make_affinity(simdata.data)
aff_copy = compute.make_affinity(*simdata.data)
# generated affinity matrices are identical regardless of how args provided
assert all(np.allclose(a1, a2) for (a1, a2) in zip(aff, aff_copy))
# outputs are square with shape (samples, samples)
assert all(a.shape == (len(d), len(d)) for a, d in zip(aff, simdata.data))
# all outputs are entirely positive (i.e., similarity / affinity)
assert all(np.all(a > 0) for a in aff)
return aff
def clusteringS(self, S_X):
"""Apply Spectral Clustering on Stable data.
SpectralClustering is performed on the S class instances. The number of clusters
is specified by the global attribute num_S. At the end of the step, the original
labels will be discarded and the new labels will be attached to the end of the data
for convenience. The assign_labels for the SpectralClustering is "discrete" and
random_state is set to 0.
Arguments
---------
S_X: list
The progressor data.
Returns
-------
S_X_clustered: numpy.ndarray
The combined clustered progressor data with new labels attached to the end.
"""
affinity_networks = compute.make_affinity(S_X,
metric='euclidean',
K=self.numS,
mu=0.5)
clusteringS = SpectralClustering(affinity='precomputed',
n_clusters=self.numS,
random_state=0).fit(affinity_networks)
labels = clusteringS.labels_
labels = labels + self.numP
S_X = np.asarray(S_X)
labels = labels.reshape(-1, 1)
S_X_clustered = np.concatenate((S_X, labels), axis=1)
if self.debug_mode == True:
print("S labels", np.reshape(labels, (1, -1)))
print("S_X_clustered", S_X_clustered)
return S_X_clustered
def fuse_and_label(data, K, mu, n_clusters, metric):
"""
Generates fusion + cluster assignments for given hyperparameters
Small helper function to be used for parallelization of gridsearch
Parameters
----------
data : list of numpy.ndarray
K : int
mu : float
n_clusters : list of int
metric : str
Returns
-------
fusion : numpy.ndarray
labels : list of numpy.ndarray
"""
aff = compute.make_affinity(*data,
K=K,
mu=mu,
metric=metric,
normalize=True)
if isinstance(aff, list) and len(aff) > 1:
fusion = compute.snf(*aff, K=K)
else:
fusion = aff
labels = [
spectral_clustering(fusion, ncl, random_state=1234)
for ncl in n_clusters
]
return fusion, labels
def test_affinity_zscore():
aff = compute.make_affinity(data1)
out = metrics.affinity_zscore(aff, label, seed=1234)
assert isinstance(out, float)
def test_silhouette_score():
aff = compute.make_affinity(data1)
out = metrics.silhouette_score(aff, label)
assert isinstance(out, float)
def test_silhouette_samples():
aff = compute.make_affinity(data1)
out = metrics._silhouette_samples(aff, label)
assert out.shape == label.shape
with pytest.raises(ValueError):
metrics._silhouette_samples(aff, np.ones(len(aff)))
def test_rank_feature_by_nmi():
aff = compute.make_affinity(data1, data2)
out = compute.snf(*aff)
inp = [(data, 'sqeuclidean') for data in [data1, data2]]
metrics.rank_feature_by_nmi(inp, out)