def __call__(self, original, sample_rate):
"""Augment original waveform
Parameters
----------
original : `np.ndarray`
(n_samples, n_channels) waveform.
sample_rate : `int`
Sample rate.
Returns
-------
augmented : `np.ndarray`
(n_samples, n_channels) noise-augmented waveform.
"""
raw_audio = RawAudio(sample_rate=sample_rate, mono=True)
# accumulate enough noise to cover duration of original waveform
noises = []
len_left = len(original)
while len_left > 0:
# select noise file at random
file = random.choice(self.files_)
# select noise segment at random
segment = next(random_segment(file['gaps'], weighted=False))
duration = segment.duration
segment_len = duration * sample_rate
# if noise segment is longer than what is needed, crop it at random
if segment_len > len_left:
duration = len_left / sample_rate
segment = next(random_subsegment(segment, duration))
noise = raw_audio.crop(file,
segment,
mode='center',
fixed=duration)
# decrease the `len_left` value by the size of the returned noise
len_left -= len(noise)
noise = normalize(noise)
noises.append(noise)
# concatenate
# FIXME: use fade-in between concatenated noises
noise = np.vstack(noises)
# select SNR at random
snr = (self.snr_max -
self.snr_min) * np.random.random_sample() + self.snr_min
alpha = np.exp(-np.log(10) * snr / 20)
return normalize(original) + alpha * noise
def __call__(self, original, sample_rate):
"""Augment original waveform
Parameters
----------
original : `np.ndarray`
(n_samples, n_channels) waveform.
sample_rate : `int`
Sample rate.
Returns
-------
augmented : `np.ndarray`
(n_samples, n_channels) noise-augmented waveform.
"""
raw_audio = RawAudio(sample_rate=sample_rate, mono=True)
original_duration = len(original) / sample_rate
# accumulate enough noise to cover duration of original waveform
noises = []
left = original_duration
while left > 0:
# select noise file at random
file = random.choice(self.files_)
duration = file['duration']
# if noise file is longer than what is needed, crop it
if duration > left:
segment = next(random_subsegment(Segment(0, duration), left))
noise = raw_audio.crop(file,
segment,
mode='center',
fixed=left)
left = 0
# otherwise, take the whole file
else:
noise = raw_audio(file).data
left -= duration
noise = normalize(noise)
noises.append(noise)
# concatenate
# FIXME: use fade-in between concatenated noises
noise = np.vstack(noises)
# select SNR at random
snr = (self.snr_max -
self.snr_min) * np.random.random_sample() + self.snr_min
alpha = np.exp(-np.log(10) * snr / 20)
return normalize(original) + alpha * noise
def __call__(self, n_samples: int, sample_rate: int) -> np.ndarray:
"""Generate noise
Parameters
----------
n_samples : int
sample_rate : int
Returns
-------
noise : (n_samples, 1) np.ndarray
"""
target_duration = n_samples / sample_rate
raw_audio = RawAudio(sample_rate=sample_rate, mono=True)
# accumulate enough noise to cover duration
noises = []
left = target_duration
while left > 0:
# select noise file at random
file = random.choice(self.files_)
duration = file['duration']
# if noise file is longer than what is needed, crop it
if duration > left:
segment = next(random_subsegment(Segment(0, duration), left))
noise = raw_audio.crop(file,
segment,
mode='center',
fixed=left)
left = 0
# otherwise, take the whole file
else:
noise = raw_audio(file).data
left -= duration
noise = normalize(noise)
noises.append(noise)
# concatenate
noise = np.vstack(noises)
# TODO: use fade-in between concatenated noises
return noise
def _random_samples(self):
"""Random samples
Returns
-------
samples : generator
Generator that yields {'X': ..., 'y': ...} samples indefinitely.
"""
uris = list(self.data_)
durations = np.array([self.data_[uri]['duration'] for uri in uris])
probabilities = durations / np.sum(durations)
while True:
# choose file at random with probability
# proportional to its (annotated) duration
uri = uris[np.random.choice(len(uris), p=probabilities)]
datum = self.data_[uri]
current_file = datum['current_file']
# choose one segment at random with probability
# proportional to its duration
segment = next(random_segment(datum['segments'], weighted=True))
# choose fixed-duration subsegment at random
subsegment = next(random_subsegment(segment, self.duration))
X = self.feature_extraction.crop(current_file,
subsegment,
mode='center',
fixed=self.duration)
y = self.crop_y(datum['y'],
subsegment)
sample = {'X': X, 'y': y}
if self.mask is not None:
mask = self.crop_y(current_file[self.mask],
subsegment)
sample['mask'] = mask
for key, classes in self.file_labels_.items():
sample[key] = classes.index(current_file[key])
yield sample
def overlap_samples(self):
"""Random overlap samples
Returns
-------
samples : generator
Generator that yields {'waveform': ..., 'y': ...} samples
indefinitely.
"""
uris = list(self.data_)
durations = np.array([self.data_[uri]['duration'] for uri in uris])
probabilities = durations / np.sum(durations)
while True:
# choose file at random with probability
# proportional to its (annotated) duration
uri = uris[np.random.choice(len(uris), p=probabilities)]
datum = self.data_[uri]
current_file = datum['current_file']
# choose one segment at random with probability
# proportional to its duration
segment = next(random_segment(datum['segments'], weighted=True))
# choose random subsegment
# duration = np.random.rand() * self.duration
sequence = next(random_subsegment(segment, self.duration))
# get corresponding waveform
X = self.raw_audio_.crop(current_file,
sequence,
mode='center',
fixed=self.duration)
# get corresponding labels
y = datum['y'].crop(sequence,
mode=self.alignment,
fixed=self.duration)
yield {'waveform': normalize(X),
'y': y}
def samples(self):
labels = list(self.data_)
# batch_counter counts samples in current batch.
# as soon as it reaches batch_size, a new random duration is selected
# so that the next batch will use a different chunk duration
batch_counter = 0
batch_size = self.batch_size
batch_duration = self.min_duration + np.random.rand() * (
self.duration - self.min_duration)
while True:
# shuffle labels
np.random.shuffle(labels)
# loop on each label
for label in labels:
# load data for this label
# segment_generators, durations, files, features = \
# zip(*self.data_[label])
segments, durations, files = zip(*self.data_[label])
# choose 'per_label' files at random with probability
# proportional to the total duration of 'label' in those files
probabilities = durations / np.sum(durations)
chosen = np.random.choice(len(files),
size=self.per_label,
p=probabilities)
# loop on (randomly) chosen files
for i in chosen:
# choose one segment at random with
# probability proportional to duration
# segment = next(segment_generators[i])
segment = next(
random_segment(segments[i], weighted=self.weighted_))
# choose per_turn chunk(s) at random
for chunk in itertools.islice(
random_subsegment(segment, batch_duration),
self.per_turn):
yield {
'X':
self.feature_extraction.crop(files[i],
chunk,
mode='center',
fixed=batch_duration),
'y':
self.segment_labels_.index(label),
}
# increment number of samples in current batch
batch_counter += 1
# as soon as the batch is complete, a new random
# duration is selected so that the next batch will use
# a different chunk duration
if batch_counter == batch_size:
batch_counter = 0
batch_duration = self.min_duration + np.random.rand(
) * (self.duration - self.min_duration)
