def _get_details(self, reference, hypothesis, **kwargs):
detail = self._init_details()
matrix = get_cooccurrence_matrix(reference, hypothesis)
duration = np.sum(matrix.M)
rduration = np.sum(matrix.M, axis=1)
hduration = np.sum(matrix.M, axis=0)
# Reference entropy and reference/hypothesis cross-entropy
cross_entropy = 0.
entropy = 0.
for i, ilabel in enumerate(matrix.iter_ilabels()):
ratio = rduration[i] / duration
if ratio > 0:
entropy -= ratio * np.log(ratio)
for j, jlabel in enumerate(matrix.iter_jlabels()):
coduration = matrix[ilabel, jlabel]
if coduration > 0:
cross_entropy -= (coduration / duration) * \
np.log(coduration / hduration[j])
detail[HOMOGENEITY_CROSS_ENTROPY] = cross_entropy
detail[HOMOGENEITY_ENTROPY] = entropy
return detail
def _get_details(self, reference, hypothesis, **kwargs):
detail = self._init_details()
if not self.detection_error:
reference = reference.crop(hypothesis.get_timeline(),
mode='intersection')
hypothesis = hypothesis.crop(reference.get_timeline(),
mode='intersection')
matrix = get_cooccurrence_matrix(reference, hypothesis)
if self.per_cluster:
# biggest class in each cluster
detail[PURITY_CORRECT] = \
np.sum([matrix[L, K] / hypothesis.label_duration(K)
for K, L in matrix.argmax(axis=0).iteritems()])
# number of clusters (as float)
detail[PURITY_TOTAL] = float(matrix.shape[1])
else:
if np.prod(matrix.shape):
detail[PURITY_CORRECT] = np.sum(np.max(matrix.df.values, axis=0))
else:
detail[PURITY_CORRECT] = 0.
# total duration of clusters (with overlap)
detail[PURITY_TOTAL] = np.sum([hypothesis.label_duration(K)
for K in hypothesis.labels()])
return detail
