def iterate(self, inputs, targets):
assert len(inputs) == len(targets)
seg_count = 0
for input in inputs:
seg_count += extract_spectrogram(
input[0], settings.ONE_SEC, settings.FREQ_ELEMENTS
).shape[
1] / settings.ONE_SEC + 1 # todo: make the segments overlapping by half a second
speaker_idx = 0
segment_pos = 0
iterations = seg_count - self.batchsize + 1
if seg_count < self.batchsize:
iterations = seg_count
for start_idx in range(0, iterations, self.batchsize):
Xb = np.zeros(
(self.batchsize, 1, settings.FREQ_ELEMENTS, settings.ONE_SEC),
dtype=np.float32)
yb = []
spect = extract_spectrogram(inputs[speaker_idx,
0], settings.ONE_SEC,
settings.FREQ_ELEMENTS)
for batch_pos in range(0, self.batchsize):
if targets is not None:
yb.append(targets[speaker_idx])
if segment_pos == spect.shape[1] / settings.ONE_SEC:
# add the last segment backwards to ensure nothing is left out
Xb[batch_pos, 0] = spect[:, spect.shape[1] -
settings.ONE_SEC:spect.shape[1]]
else:
seg_idx = segment_pos * settings.ONE_SEC
Xb[batch_pos,
0] = spect[:, seg_idx:seg_idx + settings.ONE_SEC]
segment_pos += 1
if segment_pos == spect.shape[
1] / settings.ONE_SEC + 1 and speaker_idx + 1 < len(
inputs):
speaker_idx += 1
spect = extract_spectrogram(inputs[speaker_idx,
0], settings.ONE_SEC,
settings.FREQ_ELEMENTS)
segment_pos = 0
elif segment_pos == spect.shape[
1] / settings.ONE_SEC + 1 and speaker_idx + 1 == len(
inputs):
break
yield Xb[0:len(yb)], np.asarray(yb, dtype=np.int32)
def iterate(self, inputs, targets):
assert len(inputs) == len(targets)
for i in range(0, self.batches_per_epoch):
Xb = np.zeros(
(self.batchsize, 1, settings.FREQ_ELEMENTS, settings.ONE_SEC),
dtype=np.float32)
yb = np.zeros(self.batchsize, dtype=np.int32)
# here one batch is generated
for j in range(
0, self.batchsize,
self.speaker_offset_limit * self.segments_per_sentence):
speaker_idx = randint(0, len(inputs) - 1)
for speaker_offset in range(self.speaker_offset_limit):
speaker_idx_offset = speaker_idx + speaker_offset
if speaker_idx_offset >= len(inputs):
speaker_idx_offset = speaker_idx - speaker_offset
spect = extract_spectrogram(inputs[speaker_idx,
0], settings.ONE_SEC,
settings.FREQ_ELEMENTS)
self.extract_segments(Xb, yb, targets,
j + speaker_offset * 2,
speaker_idx_offset, spect,
self.segments_per_sentence)
yield Xb, yb
def __iter__(self):
bs = self.batch_size
# build as much batches as fit into the training set
for i in range((self.n_samples + bs - 1) // bs):
Xb = np.zeros((bs, 1, settings.FREQ_ELEMENTS, settings.ONE_SEC),
dtype=np.float32)
yb = np.zeros(bs, dtype=np.int32)
# here one batch is generated
for j in range(0, bs):
speaker_idx = randint(0, len(self.X) - 1)
if self.y is not None:
yb[j] = self.y[speaker_idx]
spectrogramm = extract_spectrogram(self.X[speaker_idx,
0], settings.ONE_SEC,
settings.FREQ_ELEMENTS)
seg_idx = randint(0, spectrogramm.shape[1] - settings.ONE_SEC)
Xb[j, 0] = spectrogramm[:, seg_idx:seg_idx + settings.ONE_SEC]
yield Xb, yb
def generate_cluster_data(X, y, overlapping=False):
X_cluster = np.zeros((10000, 1, settings.FREQ_ELEMENTS, settings.ONE_SEC), dtype=np.float32)
y_cluster = []
step = settings.ONE_SEC
if overlapping:
step = settings.ONE_SEC / 2
pos = 0
for i in range(len(X)):
spect = extract_spectrogram(X[i, 0], settings.ONE_SEC, settings.FREQ_ELEMENTS)
for j in range(int(spect.shape[1] / step)):
y_cluster.append(y[i])
seg_idx = j * step
try:
X_cluster[pos, 0] = spect[:, seg_idx:seg_idx + settings.ONE_SEC]
except ValueError:
# if the last segment doesn't match ignore it
pass
pos += 1
return X_cluster[0:len(y_cluster)], np.asarray(y_cluster, dtype=np.int32)
def extract(spectrogram, segment_size):
return extract_spectrogram(spectrogram, segment_size, settings.FREQ_ELEMENTS)
def extract_spectrogram(self, spectrogram):
return extract_spectrogram(spectrogram, self.duration,
self.frequency_elements)