def testMixSequences(self):
sample_rate = 10
sequence1 = music_pb2.NoteSequence()
sequence1.notes.add(pitch=60, start_time=0.5, end_time=1.0, velocity=90)
sequence1.notes.add(pitch=62, start_time=1.0, end_time=2.0, velocity=90)
sequence1.total_time = 2.0
samples1 = np.linspace(0, 1, int(sample_rate * sequence1.total_time))
sequence2 = music_pb2.NoteSequence()
sequence2.notes.add(pitch=64, start_time=0.5, end_time=1.0, velocity=90)
sequence2.total_time = 1.0
samples2 = np.linspace(0, 1, int(sample_rate * sequence2.total_time))
mixed_samples, mixed_sequence = audio_label_data_utils.mix_sequences(
[samples1, samples2], sample_rate, [sequence1, sequence2])
expected_sequence = music_pb2.NoteSequence()
expected_sequence.ticks_per_quarter = constants.STANDARD_PPQ
expected_sequence.notes.add(
pitch=60, start_time=0.5, end_time=1.0, velocity=127)
expected_sequence.notes.add(
pitch=62, start_time=1.0, end_time=2.0, velocity=127)
expected_sequence.notes.add(
pitch=64, start_time=0.5, end_time=1.0, velocity=127)
expected_sequence.notes.add(
pitch=64, start_time=1.5, end_time=2.0, velocity=127)
expected_sequence.total_time = 2.0
self.assertProtoEquals(expected_sequence, mixed_sequence)
expected_samples = np.concatenate([samples2, samples2]) * .5 + samples1 * .5
np.testing.assert_array_equal(expected_samples, mixed_samples)
def testMixSequencesTotalTime(self):
sample_rate = 10
sequence1 = music_pb2.NoteSequence()
sequence1.notes.add(pitch=60,
start_time=0.5,
end_time=1.0,
velocity=90)
sequence1.notes.add(pitch=62,
start_time=1.0,
end_time=1.5,
velocity=90)
sequence1.total_time = 1.5
samples1 = np.linspace(0, 1, int(sample_rate * 2))
sequence2 = music_pb2.NoteSequence()
sequence2.notes.add(pitch=64,
start_time=0.5,
end_time=0.9,
velocity=90)
sequence2.total_time = 0.9
samples2 = np.linspace(0, 1, int(sample_rate * 1))
mixed_samples, mixed_sequence = audio_label_data_utils.mix_sequences(
[samples1, samples2], sample_rate, [sequence1, sequence2])
expected_sequence = music_pb2.NoteSequence()
expected_sequence.ticks_per_quarter = constants.STANDARD_PPQ
expected_sequence.notes.add(pitch=60,
start_time=0.5,
end_time=1.0,
velocity=127)
expected_sequence.notes.add(pitch=62,
start_time=1.0,
end_time=1.5,
velocity=127)
expected_sequence.notes.add(pitch=64,
start_time=0.5,
end_time=0.9,
velocity=127)
expected_sequence.notes.add(pitch=64,
start_time=1.5,
end_time=1.9,
velocity=127)
# Expected time is 1.9 because the sequences are repeated according to the
# length of their associated audio. So sequence1 is not repeated at all
# (audio is 2 seconds) and sequence2 is repeated once after shifting all the
# notes by the audio length of 1 second. The final total_time is left as is
# after the last repeat, so it ends up being 1 + .9 seconds.
expected_sequence.total_time = 1.9
self.assertProtoEquals(expected_sequence, mixed_sequence)
expected_samples = np.concatenate([samples2, samples2
]) * .5 + samples1 * .5
np.testing.assert_array_equal(expected_samples, mixed_samples)
def testMixSequencesWithSustain(self):
sample_rate = 10
sequence1 = music_pb2.NoteSequence()
sequence1.notes.add(pitch=60,
start_time=0.5,
end_time=0.6,
velocity=90)
sequence1.notes.add(pitch=62,
start_time=1.0,
end_time=2.0,
velocity=90)
sequence1.total_time = 2.0
testing_lib.add_control_changes_to_sequence(sequence1, 0,
[(0.0, 64, 127),
(1.0, 64, 0)])
samples1 = np.linspace(0, 1, sample_rate * sequence1.total_time)
sequence2 = music_pb2.NoteSequence()
sequence2.notes.add(pitch=64,
start_time=0.5,
end_time=0.6,
velocity=90)
sequence2.total_time = 1.0
testing_lib.add_control_changes_to_sequence(sequence2, 0,
[(0.0, 64, 127),
(0.9, 64, 0)])
samples2 = np.linspace(0, 1, sample_rate * sequence2.total_time)
mixed_samples, mixed_sequence = audio_label_data_utils.mix_sequences(
[samples1, samples2], sample_rate, [sequence1, sequence2])
expected_sequence = music_pb2.NoteSequence()
expected_sequence.ticks_per_quarter = constants.STANDARD_PPQ
expected_sequence.notes.add(pitch=60,
start_time=0.5,
end_time=1.0,
velocity=90)
expected_sequence.notes.add(pitch=62,
start_time=1.0,
end_time=2.0,
velocity=90)
expected_sequence.notes.add(pitch=64,
start_time=0.5,
end_time=0.9,
velocity=90)
expected_sequence.notes.add(pitch=64,
start_time=1.5,
end_time=1.9,
velocity=90)
expected_sequence.total_time = 2.0
self.assertProtoEquals(expected_sequence, mixed_sequence)
expected_samples = np.concatenate([samples2, samples2
]) * .5 + samples1 * .5
np.testing.assert_array_equal(expected_samples, mixed_samples)
def testMixSequencesLongerNoteSequence(self):
sample_rate = 10
sequence1 = music_pb2.NoteSequence()
sequence1.notes.add(pitch=60,
start_time=0.5,
end_time=1.0,
velocity=90)
sequence1.notes.add(pitch=62,
start_time=1.0,
end_time=2.0,
velocity=90)
sequence1.total_time = 2.0
# samples1 will be .1 seconds shorter than sequence1
samples1 = np.linspace(0, 1,
int(sample_rate * (sequence1.total_time - .1)))
sequence2 = music_pb2.NoteSequence()
sequence2.notes.add(pitch=64,
start_time=0.5,
end_time=1.0,
velocity=90)
sequence2.total_time = 1.0
samples2 = np.linspace(0, 1, int(sample_rate * sequence2.total_time))
mixed_samples, mixed_sequence = audio_label_data_utils.mix_sequences(
[samples1, samples2], sample_rate, [sequence1, sequence2])
expected_sequence = music_pb2.NoteSequence()
expected_sequence.ticks_per_quarter = constants.STANDARD_PPQ
expected_sequence.notes.add(pitch=60,
start_time=0.5,
end_time=1.0,
velocity=127)
expected_sequence.notes.add(pitch=62,
start_time=1.0,
end_time=2.0,
velocity=127)
expected_sequence.notes.add(pitch=64,
start_time=0.5,
end_time=1.0,
velocity=127)
expected_sequence.notes.add(pitch=64,
start_time=1.5,
end_time=2.0,
velocity=127)
expected_sequence.total_time = 2.0
self.assertProtoEquals(expected_sequence, mixed_sequence)
# We expect samples1 to have 2 samples of padding and samples2 to be
# repeated 1 time fully and once with a single sample.
expected_samples = (
np.concatenate([samples2, samples2, [samples2[0]]]) * .5 +
np.concatenate([samples1, [0, 0]]) * .5)
np.testing.assert_array_equal(expected_samples, mixed_samples)
def mix_examples(mixid_exs, sample_rate, load_audio_with_librosa):
"""Mix several Examples together to create a new example."""
mixid, exs = mixid_exs
del mixid
example_samples = []
example_sequences = []
for ex_str in exs:
ex = tf.train.Example.FromString(ex_str)
wav_data = ex.features.feature['audio'].bytes_list.value[0]
if load_audio_with_librosa:
samples = audio_io.wav_data_to_samples_librosa(
wav_data, sample_rate)
else:
samples = audio_io.wav_data_to_samples(wav_data, sample_rate)
example_samples.append(samples)
ns = music_pb2.NoteSequence.FromString(
ex.features.feature['sequence'].bytes_list.value[0])
example_sequences.append(ns)
mixed_samples, mixed_sequence = audio_label_data_utils.mix_sequences(
individual_samples=example_samples,
sample_rate=sample_rate,
individual_sequences=example_sequences)
mixed_wav_data = audio_io.samples_to_wav_data(mixed_samples, sample_rate)
mixed_id = '::'.join(['mixed'] + [ns.id for ns in example_sequences])
mixed_sequence.id = mixed_id
mixed_filename = '::'.join(['mixed'] +
[ns.filename for ns in example_sequences])
mixed_sequence.filename = mixed_filename
examples = list(
audio_label_data_utils.process_record(mixed_wav_data,
mixed_sequence,
mixed_id,
min_length=0,
max_length=-1,
sample_rate=sample_rate))
assert len(examples) == 1
return examples[0]