def call(self, z_pitch, z_vel, z=None):
"""Forward pass for the MIDI decoder.
Args:
z_pitch: Tensor containing encoded pitch in MIDI scale. [batch, time, 1].
z_vel: Tensor containing encoded velocity in MIDI scale. [batch, time, 1].
z: Additional non-MIDI latent tensor. [batch, time, n_z]
Returns:
A dictionary to feed into a processor group.
"""
# pylint: disable=unused-argument
# x = tf.concat([z_pitch, z_vel], axis=-1) # TODO(jesse): Allow velocity.
x = z_pitch
x = self.net(x) if z is None else self.net([x, z])
if self.norm is not None:
x = self.norm(x)
x = self.dense_out(x)
outputs = nn.split_to_dict(x, self.output_splits)
if self.f0_residual:
outputs['f0_midi'] += z_pitch
outputs['f0_hz'] = core.midi_to_hz(outputs['f0_midi'])
return outputs
def test_midi_to_hz_is_accurate(self):
"""Tests converting between MIDI values and their frequencies in hertz."""
midi = np.arange(128)
librosa_hz = librosa.midi_to_hz(midi)
with self.cached_session() as sess:
tf_hz = sess.run(core.midi_to_hz(midi))
self.assertAllClose(librosa_hz, tf_hz)
def _add_f0_from_midi(ex, midi_notes):
"""Add fundamental frequency (f0) estimate."""
beam.metrics.Metrics.counter('prepare-tfrecord', 'compute-f0').inc()
# TODO: hardcoded, but at some point should be read from a file
f0_hz_midi = core.midi_to_hz(midi_notes).numpy()
ex = dict(ex)
ex.update({'f0_hz_midi': f0_hz_midi.astype(np.float32)})
return ex
def _add_f0(ex):
"""Add fundamental frequency (f0) estimate."""
beam.metrics.Metrics.counter('prepare-tfrecord', 'compute-f0').inc()
# TODO: hardcoded, but at some point should be read from a file
f0_hz = core.midi_to_hz([48, 52, 55, 60, 64, 67, 72, 76, 79, 84, 88,
91]).numpy()
ex = dict(ex)
ex.update({'f0_hz': f0_hz.astype(np.float32)})
return ex
def _un_processing(self, features):
# features['f0_scaled'] = hz_to_midi(features['f0_hz']) / F0_RANGE
# to scale, 1. hztomidi, 2. divide by f0range
# to unscale, 1. * f0range, 2. miditohz
features['f0_hz'] = midi_to_hz(features['f0_scaled'] * F0_RANGE)
# features['ld_scaled'] = (features['loudness_db'] / LD_RANGE) + 1.0
# to scale, 1. / ldrange, 2. + 1
# to unscale, 1. -1, 2. * ldrange
features['loudness_db'] = (features['ld_scaled'] - 1) * LD_RANGE
return features
def test_harmonic_to_sinusoidal(self): f0_hz = core.midi_to_hz([80, 81, 82, 81, 80])[np.newaxis, :, np.newaxis] harm_amps = np.ones(shape=(1, 5, 3)) harm_amps /= np.sum(harm_amps, axis=-1, keepdims=True) amps, sin_freqs = core.harmonic_to_sinusoidal(10, harm_amps, f0_hz) sin_freqs = np.squeeze(sin_freqs) f0_hz = np.squeeze(f0_hz) self.assertAllClose(amps, harm_amps * 10) self.assertAllClose(sin_freqs[..., 0], f0_hz) self.assertAllClose(sin_freqs[..., 1], f0_hz * 2) self.assertAllClose(sin_freqs[..., 2], f0_hz * 3)
def test_midi_to_hz_is_accurate(self):
"""Tests converting between MIDI values and their frequencies in hertz."""
midi = np.arange(128)
librosa_hz = librosa.midi_to_hz(midi)
tf_hz = core.midi_to_hz(midi)
self.assertAllClose(librosa_hz, tf_hz)
