def apply(self, wav):
"""Computes distance between sliding windows embeddings
Parameter
---------
wav : str
Path to wav audio file
Returns
-------
predictions : SlidingWindowFeature
"""
from pyannote.algorithms.stats.gaussian import Gaussian
current_file = {'uri': wav, 'medium': {'wav': wav}}
t, left, right = next(self.from_file(current_file))
y = []
for xL, xR in zip(left, right):
gL = Gaussian(covariance_type='diag').fit(xL)
gR = Gaussian(covariance_type='diag').fit(xR)
y.append(gL.divergence(gR))
y = np.array(y)
window = SlidingWindow(duration=2 * self.duration,
step=self.step,
start=0.)
return SlidingWindowFeature(y, window)
def apply(self, current_file):
"""Computes BIC distance between sliding windows
Parameter
---------
current_file : dict
Returns
-------
predictions : SlidingWindowFeature
"""
from pyannote.algorithms.stats.gaussian import Gaussian
t, left, right = next(self.from_file(current_file))
y = []
for xL, xR in zip(left, right):
gL = Gaussian(covariance_type=self.covariance_type).fit(xL)
gR = Gaussian(covariance_type=self.covariance_type).fit(xR)
y.append(gL.bic(gR, penalty_coef=0)[0])
y = np.array(y)
window = SlidingWindow(duration=2 * self.duration,
step=self.step,
start=0.)
return SlidingWindowFeature(y, window)
def get_model(
self, cluster, annotation=None, feature=None, **kwargs
):
timeline = annotation.label_timeline(cluster)
data = feature.crop(timeline)
gaussian = Gaussian(covariance_type=self.covariance_type)
gaussian.fit(data)
return gaussian
def __call__(self, starting_point, features=None):
current_gaussian = None
current_label = None
current_segment = None
copy = starting_point.copy()
for segment, track, label in starting_point.itertracks(label=True):
data = features.crop(segment)
gaussian = Gaussian(covariance_type=self.covariance_type)
gaussian.fit(data)
if current_gaussian is None:
current_gaussian = gaussian
current_segment = segment
current_label = starting_point[segment, track]
continue
gap = (current_segment ^ segment).duration
current_segment = segment
# stop merging if gap is large
if gap > self.max_gap:
current_gaussian = gaussian
current_label = label
continue
delta_bic, merged_gaussian = current_gaussian.bic(
gaussian, penalty_coef=self.penalty_coef)
# stop merging if similariy is small
if delta_bic < 0.0:
current_gaussian = gaussian
current_label = label
continue
# merge in any other situations
TEMPLATE = (
"Merging {cluster1} and {cluster2} with "
"(BIC = {bic:g})."
)
message = TEMPLATE.format(
cluster1=current_label,
cluster2=label,
bic=delta_bic)
self.logger.debug(message)
current_gaussian = merged_gaussian
copy[segment, track] = current_label
return copy
def __call__(self, starting_point, features=None):
current_gaussian = None
current_label = None
current_segment = None
copy = starting_point.copy()
for segment, track, label in starting_point.itertracks(label=True):
data = features.crop(segment)
gaussian = Gaussian(covariance_type=self.covariance_type)
gaussian.fit(data)
if current_gaussian is None:
current_gaussian = gaussian
current_segment = segment
current_label = starting_point[segment, track]
continue
gap = (current_segment ^ segment).duration
current_segment = segment
# stop merging if gap is large
if gap > self.max_gap:
current_gaussian = gaussian
current_label = label
continue
delta_bic, merged_gaussian = current_gaussian.bic(
gaussian, penalty_coef=self.penalty_coef)
# stop merging if similariy is small
if delta_bic < 0.0:
current_gaussian = gaussian
current_label = label
continue
# merge in any other situations
TEMPLATE = ("Merging {cluster1} and {cluster2} with "
"(BIC = {bic:g}).")
message = TEMPLATE.format(cluster1=current_label,
cluster2=label,
bic=delta_bic)
self.logger.debug(message)
current_gaussian = merged_gaussian
copy[segment, track] = current_label
return copy
def compute_model(self, cluster, parent=None):
timeline = parent.current_state.label_timeline(cluster)
data = parent.features.crop(timeline)
gaussian = Gaussian(covariance_type=self.covariance_type)
gaussian.fit(data)
return gaussian
def compare(dataset, medium_template, config_yml, output_dir):
import itertools
from pyannote.algorithms.stats.gaussian import Gaussian
# load configuration file
with open(config_yml, 'r') as fp:
config = yaml.load(fp)
X, y_true = generate_test(dataset, medium_template, config)
n_sequences = X.shape[0]
gaussians = []
for x in X:
g = Gaussian(covariance_type='diag').fit(x)
gaussians.append(g)
bic = np.zeros((n_sequences, n_sequences), dtype=np.float)
for i, j in itertools.combinations(range(n_sequences), 2):
bic[i, j], _ = gaussians[i].bic(gaussians[j], penalty_coef=0.)
distances = squareform(bic, checks=False)
# -- distances distributions
plot_distributions(y_true,
distances,
output_dir + '/plot.bic',
xlim=(0, 20),
ymax=0.5,
nbins=100)
# -- precision / recall curve
auc = plot_precision_recall_curve(y_true, -distances,
output_dir + '/plot.bic')
msg = 'BIC | AUC = {auc:.2f}%'
print(msg.format(auc=100 * auc))
# -- det curve
eer = plot_det_curve(y_true, -distances, output_dir + '/plot.bic')
msg = 'BIC | EER = {eer:.2f}%'
print(msg.format(eer=100 * eer))
divergence = np.zeros((n_sequences, n_sequences), dtype=np.float)
for i, j in itertools.combinations(range(n_sequences), 2):
divergence[i, j] = gaussians[i].divergence(gaussians[j])
distances = squareform(divergence, checks=False)
# -- distances distributions
plot_distributions(y_true,
distances,
output_dir + '/plot.divergence',
xlim=(0, 20),
ymax=0.5,
nbins=100)
# -- precision / recall curve
auc = plot_precision_recall_curve(y_true, -distances,
output_dir + '/plot.divergence')
msg = 'Divergence | AUC = {auc:.2f}%'
print(msg.format(auc=100 * auc))
# -- det curve
eer = plot_det_curve(y_true, -distances, output_dir + '/plot.divergence')
msg = 'Divergence | EER = {eer:.2f}%'
print(msg.format(eer=100 * eer))
def compute_model(self, cluster, parent=None):
timeline = parent.current_state.label_timeline(cluster)
data = parent.features.crop(timeline)
gaussian = Gaussian(covariance_type=self.covariance_type)
gaussian.fit(data)
return gaussian
# randomly select (at most) 100 sequences from each speaker to ensure
# all speakers have the same importance in the evaluation
unique, y, counts = np.unique(y, return_inverse=True, return_counts=True)
n_speakers = len(unique)
indices = []
for speaker in range(n_speakers):
i = np.random.choice(np.where(y == speaker)[0],
size=min(100, counts[speaker]),
replace=False)
indices.append(i)
indices = np.hstack(indices)
X, y = X[indices], y[indices, np.newaxis]
# one Gaussian per segment
from pyannote.algorithms.stats.gaussian import Gaussian
full = [Gaussian(covariance_type='full').fit(x) for x in X]
diag = [Gaussian(covariance_type='diag').fit(x) for x in X]
# compute BIC and divergence between every pair of sequences
import itertools
n_sequences = len(X)
bic = np.zeros((n_sequences, n_sequences), dtype=np.float32)
div = np.zeros((n_sequences, n_sequences), dtype=np.float32)
for i, j in itertools.combinations(range(n_sequences), 2):
gi, gj = full[i], full[j]
bic[i, j] = gi.bic(gj, penalty_coef=0)[0]
bic[j, i] = bic[i, j]
gi, gj = diag[i], diag[j]