def test_load(sample_list, sample_annotation, sample_scores):
coParser = CoParser(uris=sample_list,
annotation=sample_annotation,
scores=sample_scores)
for uri, annotation, scores in coParser.iter('uris', 'annotation', 'scores'):
break
assert uri == 'uri1'
assert list(annotation.itertracks(label=True)) == [
(Segment(1, 3.5), 0, 'alice'),
(Segment(3, 7.5), 1, 'barbara'),
(Segment(6, 9), 2, 'chris')]
assert list(scores.itervalues()) == [
(Segment(1, 3.5), 0, 'alice', 0.8),
(Segment(1, 3.5), 1, 'barbara', 0.1),
(Segment(1, 3.5), 2, 'chris', 0.1),
(Segment(3, 7.5), 3, 'barbara', 0.5),
(Segment(3, 7.5), 4, 'chris', 0.4),
(Segment(6, 9), 5, 'alice', 0.1),
(Segment(6, 9), 6, 'barbara', 0.2),
(Segment(6, 9), 7, 'chris', 0.7)]
def do_identification(
references_mdtm, hypothesis_mdtm, uris_lst=None, eval_uem=None
):
ier = IdentificationErrorRate()
iter_over = {
'reference': references_mdtm,
'hypothesis': hypothesis_mdtm
}
if uris_lst:
iter_over['uris'] = uris_lst
else:
iter_over['uris'] = 'reference'
if eval_uem:
iter_over['uem'] = eval_uem
coParser = CoParser(**iter_over)
for uri, ref, hyp, uem in coParser.iter('uris', 'reference', 'hypothesis',
'uem'):
rate = ier(ref, hyp, uem=uem)
sys.stdout.write('{uri:s}: {ier:3.2f}%\n'.format(uri=uri, ier=100 * rate))
sys.stdout.flush()
sys.stdout.write('Total: {ier:3.2f}%'.format(ier=100 * abs(ier)))
def test_load(sample_list, sample_annotation, sample_scores):
coParser = CoParser(uris=sample_list,
annotation=sample_annotation,
scores=sample_scores)
for uri, annotation, scores in coParser.iter('uris', 'annotation',
'scores'):
break
assert uri == 'uri1'
assert list(annotation.itertracks(label=True)) == [
(Segment(1, 3.5), 0, 'alice'), (Segment(3, 7.5), 1, 'barbara'),
(Segment(6, 9), 2, 'chris')
]
assert list(scores.itervalues()) == [(Segment(1, 3.5), 0, 'alice', 0.8),
(Segment(1, 3.5), 1, 'barbara', 0.1),
(Segment(1, 3.5), 2, 'chris', 0.1),
(Segment(3, 7.5), 3, 'barbara', 0.5),
(Segment(3, 7.5), 4, 'chris', 0.4),
(Segment(6, 9), 5, 'alice', 0.1),
(Segment(6, 9), 6, 'barbara', 0.2),
(Segment(6, 9), 7, 'chris', 0.7)]
def do_train(
uris_lst,
references_mdtm,
features_pkl,
model_pkl,
n_components=16,
covariance_type='diag',
min_duration=0.250,
):
hmm = ViterbiHMM(n_components=n_components,
covariance_type=covariance_type,
random_state=None,
thresh=1e-2,
min_covar=1e-3,
n_iter=10,
disturb=0.05,
sampling=1000,
min_duration=min_duration)
# iterate over all uris in a synchronous manner
coParser = CoParser(uris=uris_lst,
reference=references_mdtm,
features=features_pkl)
references, features = coParser.generators('reference', 'features')
hmm.fit(references, features)
with open(model_pkl, 'wb') as f:
pickle.dump(hmm, f)
def do_stats(reference_repere, hypothesis_repere,
uris_lst=None, eval_uem=None,
collar=0., unknown=False):
identificationErrorAnalysis = IdentificationErrorAnalysis(
collar=collar, unknown=unknown)
iter_over = {
'reference': reference_repere,
'hypothesis': hypothesis_repere
}
if uris_lst:
iter_over['uris'] = uris_lst
else:
iter_over['uris'] = 'reference'
if eval_uem:
iter_over['uem'] = eval_uem
coParser = CoParser(**iter_over)
for uri, ref, hyp, uem in coParser.iter(
'uris', 'reference', 'hypothesis', 'uem'
):
# perform error analysis
analysis = identificationErrorAnalysis.matrix(ref, hyp, uem=uem)
# obtain list of speakers
speakers = [speaker for speaker in analysis.get_rows()]
# loop on each speaker
for spk in speakers:
if isinstance(spk, tuple) and spk[0] == 'false alarm':
continue
ref = analysis[spk, 'reference']
hyp = analysis[spk, 'hypothesis']
cor = analysis[spk, 'correct']
conf = analysis[spk, 'confusion']
miss = analysis[spk, 'missed detection']
fa = analysis[spk, 'false alarm']
ier = (conf + miss + fa) / ref
sys.stdout.write(
TEMPLATE.format(uri=uri, spk=spk, ier=ier, ref=ref, hyp=hyp,
cor=cor, conf=conf, miss=miss, fa=fa))
sys.stdout.flush()
def do_train(
uris_lst, references_mdtm, features_pkl, model_pkl,
n_components=16, covariance_type='diag', min_duration=0.250,
):
hmm = ViterbiHMM(
n_components=n_components, covariance_type=covariance_type,
random_state=None, thresh=1e-2, min_covar=1e-3, n_iter=10,
disturb=0.05, sampling=1000, min_duration=min_duration)
# iterate over all uris in a synchronous manner
coParser = CoParser(uris=uris_lst,
reference=references_mdtm,
features=features_pkl)
references, features = coParser.generators('reference', 'features')
hmm.fit(references, features)
with open(model_pkl, 'wb') as f:
pickle.dump(hmm, f)
def do_stats(reference_repere, hypothesis_repere,
uris_lst=None, eval_uem=None,
unknown=False):
iter_over = {
'reference': reference_repere,
'hypothesis': hypothesis_repere
}
if uris_lst:
iter_over['uris'] = uris_lst
else:
iter_over['uris'] = 'reference'
if eval_uem:
iter_over['uem'] = eval_uem
coParser = CoParser(**iter_over)
identificationErrorRate = IdentificationErrorRate(unknown=unknown)
for uri, ref, hyp, uem in coParser.iter(
'uris', 'reference', 'hypothesis', 'uem'
):
ref = ref.crop(uem, mode='intersection')
hyp = hyp.crop(uem, mode='intersection')
overlap = {}
# starts by finding overlap regions in reference
# this will iterate over all pairs of intersecting segments
for (segment, track), (segment_, track_) in ref.co_iter(ref):
# do not count twice
if (segment, track) >= (segment, track_):
continue
# overlap duration
intersection = (segment & segment_).duration
# increment overlap count
overlap[segment, track] = \
overlap.get((segment, track), 0.) + intersection
overlap[segment_, track_] = \
overlap.get((segment_, track_), 0.) + intersection
for segment, _, spk in ref.itertracks(label=True):
if isinstance(spk, Unknown):
continue
r = Annotation(uri=ref.uri, modality=ref.modality)
r[segment] = spk
h = hyp.crop(segment)
details = identificationErrorRate(r, h, detailed=True)
ovr = overlap.get((segment, track), 0)
ier = details[IER_NAME]
cor = details[IER_CORRECT]
conf = details[IER_CONFUSION] - 0.5 * ovr
miss = details[IER_MISS] + .5 * ovr
fa = details[IER_FALSE_ALARM]
sys.stdout.write(
TEMPLATE.format(uri=uri, start=segment.start, end=segment.end,
spk=spk, ier=ier, cor=cor, conf=conf, miss=miss,
fa=fa))
sys.stdout.flush()