def cal_dcc_scores(self, method='spearman', mean=True):
r""" Same as D.C.I but use correlation matrix instead of importance matrix
"""
train, test = self.create_correlation_matrix(mean=mean,
method=method,
decode=False)
train = np.abs(train)
test = np.abs(test)
d = (metrics.disentanglement_score(train) +
metrics.disentanglement_score(test)) / 2.
c = (metrics.completeness_score(train) +
metrics.completeness_score(test)) / 2.
return d, c
def cal_dcmi_scores(self, mean=True, n_neighbors=3):
r""" The same method is used for D.C.I scores, however, this metrics use
mutual information matrix (estimated by nearest neighbor method)
instead of importance matrix
Return:
tuple of 2 scalars:
- disentanglement score of mutual information
- completeness score of mutual information
"""
train, test = self.create_mutualinfo_matrix(mean=mean,
n_neighbors=n_neighbors)
d = (metrics.disentanglement_score(train) +
metrics.disentanglement_score(test)) / 2.
c = (metrics.completeness_score(train) +
metrics.completeness_score(test)) / 2.
return d, c
def cal_dcd_scores(self, n_samples=1000, lognorm=True, n_components=2):
r""" Same as D.C.I but use density matrix instead of importance matrix
"""
# smaller is better
train, test = self.create_divergence_matrix(n_samples=n_samples,
lognorm=lognorm,
n_components=n_components,
normalize_per_code=True,
decode=False)
# diag = np.diagflat(np.diag(density_mat))
# higher is better
train = 1. - train
test = 1 - test
d = (metrics.disentanglement_score(train) +
metrics.disentanglement_score(test)) / 2.
c = (metrics.completeness_score(train) +
metrics.completeness_score(test)) / 2.
return d, c