def _prepare_example(self, example):
"""Creates an example (anchor-positive) for instance discrimination."""
x = tf.math.l2_normalize(example["audio"], epsilon=1e-9)
waveform_a = data.extract_window(x)
mels_a = data.extract_log_mel_spectrogram(waveform_a)
frames_anchors = mels_a[Ellipsis, tf.newaxis]
waveform_p = data.extract_window(x)
waveform_p = waveform_p + (self._noise *
tf.random.normal(tf.shape(waveform_p)))
mels_p = data.extract_log_mel_spectrogram(waveform_p)
frames_positives = mels_p[Ellipsis, tf.newaxis]
return frames_anchors, frames_positives
def _prepare_standard_example(self, example, is_training):
"""Creates an example for supervised training."""
x = example["audio"]
if is_training:
x = data.extract_window(x)
x = tf.math.l2_normalize(x, epsilon=1e-9)
else:
x = tf.signal.frame(x,
frame_length=self._n_frames * 160,
frame_step=self._n_frames * 160,
pad_end=True)
x = tf.math.l2_normalize(x, axis=-1, epsilon=1e-9)
x = data.extract_log_mel_spectrogram(x)
x = x[Ellipsis, tf.newaxis]
y = example["label"]
return x, y