def test_transition_bias_unique_speaker(self):
"""Test when the input cluster_id sequences contain a unique speaker
and therefore no speaker changes"""
transition_bias, _ = utils.estimate_transition_bias(cluster_ids=[[1] *
100])
self.assertTrue(np.log(transition_bias) != -np.inf)
self.assertTrue(np.log(1 - transition_bias) != -np.inf)
def test_transition_bias_always_changing_speaker(self):
"""Test when in the input cluster_id sequences the speaker always
changes"""
transition_bias, _ = utils.estimate_transition_bias(
cluster_ids=[[1, 2, 1], [2, 1, 2]])
self.assertTrue(np.log(transition_bias) != -np.inf)
self.assertTrue(np.log(1 - transition_bias) != -np.inf)
def test_transition_bias_empty_sequences(self):
"""Test when the input cluster_id sequences are empty"""
(transition_bias, denominator) = utils.estimate_transition_bias(
cluster_ids=[[], [], []])
self.assertTrue(np.log(transition_bias) != -np.inf)
self.assertTrue(np.log(1 - transition_bias) != -np.inf)
self.assertTrue(denominator != 0)
def fit(self, train_sequences, train_cluster_ids, args):
"""Fit UISRNN model.
Args:
train_sequences: Either a list of training sequences, or a single
concatenated training sequence:
1. train_sequences is list, and each element is a 2-dim numpy array
of real numbers, of size: `length * D`.
The length varies among different sequences, but the D is the same.
In speaker diarization, each sequence is the sequence of speaker
embeddings of one utterance.
2. train_sequences is a single concatenated sequence, which is a
2-dim numpy array of real numbers. See `fit_concatenated()`
for more details.
train_cluster_ids: Ground truth labels for train_sequences:
1. if train_sequences is a list, this must also be a list of the same
size, each element being a 1-dim list or numpy array of strings.
2. if train_sequences is a single concatenated sequence, this
must also be the concatenated 1-dim list or numpy array of strings
args: Training configurations. See `arguments.py` for details.
Raises:
TypeError: If train_sequences or train_cluster_ids is of wrong type.
"""
if isinstance(train_sequences, np.ndarray):
# train_sequences is already the concatenated sequence
if self.estimate_transition_bias:
# see issue #55: https://github.com/google/uis-rnn/issues/55
self.logger.print(
2,
'Warning: transition_bias cannot be correctly estimated from a '
'concatenated sequence; train_sequences will be treated as a '
'single sequence. This can lead to inaccurate estimation of '
'transition_bias. Please, consider estimating transition_bias '
'before concatenating the sequences and passing it as argument.'
)
train_sequences = [train_sequences]
train_cluster_ids = [train_cluster_ids]
elif isinstance(train_sequences, list):
# train_sequences is a list of un-concatenated sequences
# we will concatenate it later, after estimating transition_bias
pass
else:
raise TypeError('train_sequences must be a list or numpy.ndarray')
# estimate transition_bias
if self.estimate_transition_bias:
(transition_bias, transition_bias_denominator
) = utils.estimate_transition_bias(train_cluster_ids)
# set or update transition_bias
if self.transition_bias is None:
self.transition_bias = transition_bias
self.transition_bias_denominator = transition_bias_denominator
else:
self.transition_bias = (
self.transition_bias * self.transition_bias_denominator +
transition_bias * transition_bias_denominator) / (
self.transition_bias_denominator +
transition_bias_denominator)
self.transition_bias_denominator += transition_bias_denominator
# concatenate train_sequences
(concatenated_train_sequence,
concatenated_train_cluster_id) = utils.concatenate_training_data(
train_sequences, train_cluster_ids,
args.enforce_cluster_id_uniqueness, True)
self.fit_concatenated(concatenated_train_sequence,
concatenated_train_cluster_id, args)
