def get_pipeline(config, eval_ratio):
"""Returns the Pipeline instance which creates the RNN dataset.
Args:
config: An ImprovRnnConfig object.
eval_ratio: Fraction of input to set aside for evaluation set.
Returns:
A pipeline.Pipeline instance.
"""
all_transpositions = config.transpose_to_key is None
partitioner = pipelines_common.RandomPartition(
music_pb2.NoteSequence,
['eval_lead_sheets', 'training_lead_sheets'],
[eval_ratio])
dag = {partitioner: dag_pipeline.DagInput(music_pb2.NoteSequence)}
for mode in ['eval', 'training']:
time_change_splitter = note_sequence_pipelines.TimeChangeSplitter(
name='TimeChangeSplitter_' + mode)
quantizer = note_sequence_pipelines.Quantizer(
steps_per_quarter=config.steps_per_quarter, name='Quantizer_' + mode)
lead_sheet_extractor = lead_sheet_pipelines.LeadSheetExtractor(
min_bars=7, max_steps=512, min_unique_pitches=3, gap_bars=1.0,
ignore_polyphonic_notes=False, all_transpositions=all_transpositions,
name='LeadSheetExtractor_' + mode)
encoder_pipeline = EncoderPipeline(config, name='EncoderPipeline_' + mode)
dag[time_change_splitter] = partitioner[mode + '_lead_sheets']
dag[quantizer] = time_change_splitter
dag[lead_sheet_extractor] = quantizer
dag[encoder_pipeline] = lead_sheet_extractor
dag[dag_pipeline.DagOutput(mode + '_lead_sheets')] = encoder_pipeline
return dag_pipeline.DAGPipeline(dag)
def testMelodyRNNPipeline(self):
note_sequence = magenta.common.testing_lib.parse_test_proto(
music_pb2.NoteSequence, """
time_signatures: {
numerator: 4
denominator: 4}
tempos: {
qpm: 120}""")
magenta.music.testing_lib.add_track_to_sequence(
note_sequence, 0, [(12, 100, 0.00, 2.0), (11, 55, 2.1, 5.0),
(40, 45, 5.1, 8.0), (55, 120, 8.1, 11.0),
(53, 99, 11.1, 14.1)])
quantizer = note_sequence_pipelines.Quantizer(steps_per_quarter=4)
melody_extractor = melody_pipelines.MelodyExtractor(
min_bars=7,
min_unique_pitches=5,
gap_bars=1.0,
ignore_polyphonic_notes=False)
one_hot_encoding = magenta.music.OneHotEventSequenceEncoderDecoder(
magenta.music.MelodyOneHotEncoding(self.config.min_note,
self.config.max_note))
quantized = quantizer.transform(note_sequence)[0]
melody = melody_extractor.transform(quantized)[0]
melody.squash(self.config.min_note, self.config.max_note,
self.config.transpose_to_key)
one_hot = one_hot_encoding.encode(melody)
expected_result = {'training_melodies': [one_hot], 'eval_melodies': []}
pipeline_inst = melody_rnn_pipeline.get_pipeline(self.config,
eval_ratio=0.0)
result = pipeline_inst.transform(note_sequence)
self.assertEqual(expected_result, result)
def get_pipeline(config, eval_ratio):
"""Returns the Pipeline instance which creates the RNN dataset.
Args:
config: A DrumsRnnConfig object.
eval_ratio: Fraction of input to set aside for evaluation set.
Returns:
A pipeline.Pipeline instance.
"""
partitioner = pipelines_common.RandomPartition(
music_pb2.NoteSequence,
['eval_drum_tracks', 'training_drum_tracks'],
[eval_ratio])
dag = {partitioner: dag_pipeline.DagInput(music_pb2.NoteSequence)}
for mode in ['eval', 'training']:
time_change_splitter = note_sequence_pipelines.TimeChangeSplitter(
name='TimeChangeSplitter_' + mode)
quantizer = note_sequence_pipelines.Quantizer(
steps_per_quarter=config.steps_per_quarter, name='Quantizer_' + mode)
drums_extractor = drum_pipelines.DrumsExtractor(
min_bars=7, max_steps=512, gap_bars=1.0, name='DrumsExtractor_' + mode)
encoder_pipeline = event_sequence_pipeline.EncoderPipeline(
magenta.music.DrumTrack, config.encoder_decoder,
name='EncoderPipeline_' + mode)
dag[time_change_splitter] = partitioner[mode + '_drum_tracks']
dag[quantizer] = time_change_splitter
dag[drums_extractor] = quantizer
dag[encoder_pipeline] = drums_extractor
dag[dag_pipeline.DagOutput(mode + '_drum_tracks')] = encoder_pipeline
return dag_pipeline.DAGPipeline(dag)
def testDrumsRNNPipeline(self):
note_sequence = magenta.common.testing_lib.parse_test_proto(
music_pb2.NoteSequence, """
time_signatures: {
numerator: 4
denominator: 4}
tempos: {
qpm: 120}""")
magenta.music.testing_lib.add_track_to_sequence(
note_sequence,
0, [(36, 100, 0.00, 2.0), (40, 55, 2.1, 5.0), (44, 80, 3.6, 5.0),
(41, 45, 5.1, 8.0), (64, 100, 6.6, 10.0), (55, 120, 8.1, 11.0),
(39, 110, 9.6, 9.7), (53, 99, 11.1, 14.1),
(51, 40, 12.6, 13.0), (55, 100, 14.1, 15.0),
(54, 90, 15.6, 17.0), (60, 100, 17.1, 18.0)],
is_drum=True)
quantizer = note_sequence_pipelines.Quantizer(steps_per_quarter=4)
drums_extractor = drum_pipelines.DrumsExtractor(min_bars=7,
gap_bars=1.0)
one_hot_encoding = magenta.music.OneHotEventSequenceEncoderDecoder(
magenta.music.MultiDrumOneHotEncoding())
quantized = quantizer.transform(note_sequence)[0]
drums = drums_extractor.transform(quantized)[0]
one_hot = one_hot_encoding.encode(drums)
expected_result = {
'training_drum_tracks': [one_hot],
'eval_drum_tracks': []
}
pipeline_inst = drums_rnn_pipeline.get_pipeline(self.config,
eval_ratio=0.0)
result = pipeline_inst.transform(note_sequence)
self.assertEqual(expected_result, result)
def get_pipeline(config, min_events, max_events, eval_ratio):
"""Returns the Pipeline instance which creates the RNN dataset.
Args:
config: A PerformanceRnnConfig.
min_events: Minimum number of events for an extracted sequence.
max_events: Maximum number of events for an extracted sequence.
eval_ratio: Fraction of input to set aside for evaluation set.
Returns:
A pipeline.Pipeline instance.
"""
# Stretch by -5%, -2.5%, 0%, 2.5%, and 5%.
stretch_factors = [0.95, 0.975, 1.0, 1.025, 1.05]
# Transpose no more than a major third.
transposition_range = range(-3, 4)
partitioner = pipelines_common.RandomPartition(
music_pb2.NoteSequence, ['eval_performances', 'training_performances'],
[eval_ratio])
dag = {partitioner: dag_pipeline.DagInput(music_pb2.NoteSequence)}
for mode in ['eval', 'training']:
sustain_pipeline = note_sequence_pipelines.SustainPipeline(
name='SustainPipeline_' + mode)
stretch_pipeline = note_sequence_pipelines.StretchPipeline(
stretch_factors, name='StretchPipeline_' + mode)
splitter = note_sequence_pipelines.Splitter(hop_size_seconds=30.0,
name='Splitter_' + mode)
quantizer = note_sequence_pipelines.Quantizer(
steps_per_second=config.steps_per_second, name='Quantizer_' + mode)
transposition_pipeline = note_sequence_pipelines.TranspositionPipeline(
transposition_range, name='TranspositionPipeline_' + mode)
perf_extractor = PerformanceExtractor(
min_events=min_events,
max_events=max_events,
num_velocity_bins=config.num_velocity_bins,
name='PerformanceExtractor_' + mode)
encoder_pipeline = EncoderPipeline(config,
name='EncoderPipeline_' + mode)
dag[sustain_pipeline] = partitioner[mode + '_performances']
if mode == 'eval':
# No stretching in eval.
dag[splitter] = sustain_pipeline
else:
dag[stretch_pipeline] = sustain_pipeline
dag[splitter] = stretch_pipeline
dag[quantizer] = splitter
if mode == 'eval':
# No transposition in eval.
dag[perf_extractor] = quantizer
else:
dag[transposition_pipeline] = quantizer
dag[perf_extractor] = transposition_pipeline
dag[encoder_pipeline] = perf_extractor
dag[dag_pipeline.DagOutput(mode + '_performances')] = encoder_pipeline
return dag_pipeline.DAGPipeline(dag)
def get_pipeline(config, eval_ratio):
"""Returns the Pipeline instance which creates the RNN dataset.
Args:
config: A MelodyRnnConfig object.
eval_ratio: Fraction of input to set aside for evaluation set.
Returns:
A pipeline.Pipeline instance.
"""
partitioner = pipelines_common.RandomPartition(
music_pb2.NoteSequence,
['eval_melodies', 'training_melodies'],
[eval_ratio])
dag = {partitioner: dag_pipeline.DagInput(music_pb2.NoteSequence)}
for mode in ['eval', 'training']:
time_change_splitter = note_sequence_pipelines.TimeChangeSplitter(
name='TimeChangeSplitter_' + mode)
quantizer = note_sequence_pipelines.Quantizer(
steps_per_quarter=config.steps_per_quarter, name='Quantizer_' + mode)
melody_extractor = melody_pipelines.MelodyExtractor(
min_bars=7, max_steps=512, min_unique_pitches=5,
gap_bars=1.0, ignore_polyphonic_notes=True,
name='MelodyExtractor_' + mode)
encoder_pipeline = EncoderPipeline(config, name='EncoderPipeline_' + mode)
dag[time_change_splitter] = partitioner[mode + '_melodies']
dag[quantizer] = time_change_splitter
dag[melody_extractor] = quantizer
dag[encoder_pipeline] = melody_extractor
dag[dag_pipeline.DagOutput(mode + '_melodies')] = encoder_pipeline
return dag_pipeline.DAGPipeline(dag)
def get_pipeline(config, eval_ratio=0.0):
partitioner = pipelines_common.RandomPartition(
music_pb2.NoteSequence, ['eval_melodies', 'training_melodies'],
[eval_ratio])
dag = {partitioner: dag_pipeline.DagInput(music_pb2.NoteSequence)}
for mode in ['eval', 'training']:
time_change_splitter = note_sequence_pipelines.TimeChangeSplitter(
name='TimeChangeSplitter_' + mode)
repeat_sequence = RepeatSequence(min_duration=16,
name='RepeatSequence_' + mode)
transposition_pipeline = note_sequence_pipelines.TranspositionPipeline(
(0, ), name='TranspositionPipeline_' + mode)
quantizer = note_sequence_pipelines.Quantizer(
steps_per_quarter=config.steps_per_quarter,
name='Quantizer_' + mode)
melody_extractor = melody_pipelines.MelodyExtractor(
min_bars=7,
max_steps=512,
min_unique_pitches=5,
gap_bars=1.0,
ignore_polyphonic_notes=True,
name='MelodyExtractor_' + mode)
encoder_pipeline = EncoderPipeline(config,
name='EncoderPipeline_' + mode)
dag[time_change_splitter] = partitioner[mode + '_melodies']
dag[repeat_sequence] = time_change_splitter
dag[quantizer] = repeat_sequence
dag[transposition_pipeline] = quantizer
dag[melody_extractor] = transposition_pipeline
dag[encoder_pipeline] = melody_extractor
dag[dag_pipeline.DagOutput(mode + '_melodies')] = encoder_pipeline
return dag_pipeline.DAGPipeline(dag)
def get_pipeline(config, min_steps, max_steps, eval_ratio):
"""Returns the Pipeline instance which creates the RNN dataset.
Args:
config: An EventSequenceRnnConfig.
min_steps: Minimum number of steps for an extracted sequence.
max_steps: Maximum number of steps for an extracted sequence.
eval_ratio: Fraction of input to set aside for evaluation set.
Returns:
A pipeline.Pipeline instance.
"""
# Transpose up to a major third in either direction.
# Because our current dataset is Bach chorales, transposing more than a major
# third in either direction probably doesn't makes sense (e.g., because it is
# likely to exceed normal singing range).
transposition_range = range(-4, 5)
partitioner = pipelines_common.RandomPartition(
music_pb2.NoteSequence, ['eval_poly_tracks', 'training_poly_tracks'],
[eval_ratio])
dag = {partitioner: dag_pipeline.DagInput(music_pb2.NoteSequence)}
for mode in ['eval', 'training']:
time_change_splitter = note_sequence_pipelines.TimeChangeSplitter(
name='TimeChangeSplitter_' + mode)
quantizer = note_sequence_pipelines.Quantizer(
steps_per_quarter=config.steps_per_quarter,
name='Quantizer_' + mode)
transposition_pipeline = note_sequence_pipelines.TranspositionPipeline(
transposition_range, name='TranspositionPipeline_' + mode)
poly_extractor = PolyphonicSequenceExtractor(min_steps=min_steps,
max_steps=max_steps,
name='PolyExtractor_' +
mode)
encoder_pipeline = event_sequence_pipeline.EncoderPipeline(
polyphony_lib.PolyphonicSequence,
config.encoder_decoder,
name='EncoderPipeline_' + mode)
dag[time_change_splitter] = partitioner[mode + '_poly_tracks']
dag[quantizer] = time_change_splitter
dag[transposition_pipeline] = quantizer
dag[poly_extractor] = transposition_pipeline
dag[encoder_pipeline] = poly_extractor
dag[dag_pipeline.DagOutput(mode + '_poly_tracks')] = encoder_pipeline
return dag_pipeline.DAGPipeline(dag)
def testMelodyRNNPipeline(self):
note_sequence = magenta.common.testing_lib.parse_test_proto(
note_seq.NoteSequence, """
time_signatures: {
numerator: 4
denominator: 4}
tempos: {
qpm: 120}""")
note_seq.testing_lib.add_track_to_sequence(note_sequence, 0,
[(12, 100, 0.00, 2.0),
(11, 55, 2.1, 5.0),
(40, 45, 5.1, 8.0),
(55, 120, 8.1, 11.0),
(53, 99, 11.1, 14.1)])
note_seq.testing_lib.add_chords_to_sequence(note_sequence,
[('N.C.', 0.0),
('Am9', 5.0),
('D7', 10.0)])
quantizer = note_sequence_pipelines.Quantizer(steps_per_quarter=4)
lead_sheet_extractor = lead_sheet_pipelines.LeadSheetExtractor(
min_bars=7,
min_unique_pitches=5,
gap_bars=1.0,
ignore_polyphonic_notes=False,
all_transpositions=False)
conditional_encoding = note_seq.ConditionalEventSequenceEncoderDecoder(
note_seq.OneHotEventSequenceEncoderDecoder(
note_seq.MajorMinorChordOneHotEncoding()),
note_seq.OneHotEventSequenceEncoderDecoder(
note_seq.MelodyOneHotEncoding(self.config.min_note,
self.config.max_note)))
quantized = quantizer.transform(note_sequence)[0]
lead_sheet = lead_sheet_extractor.transform(quantized)[0]
lead_sheet.squash(self.config.min_note, self.config.max_note,
self.config.transpose_to_key)
encoded = pipelines_common.make_sequence_example(
*conditional_encoding.encode(lead_sheet.chords, lead_sheet.melody))
expected_result = {
'training_lead_sheets': [encoded],
'eval_lead_sheets': []
}
pipeline_inst = improv_rnn_pipeline.get_pipeline(self.config,
eval_ratio=0.0)
result = pipeline_inst.transform(note_sequence)
self.assertEqual(expected_result, result)
def get_pipeline(config, min_steps, max_steps, eval_ratio):
"""Returns the Pipeline instance which creates the RNN dataset.
Args:
config: An EventSequenceRnnConfig.
min_steps: Minimum number of steps for an extracted sequence.
max_steps: Maximum number of steps for an extracted sequence.
eval_ratio: Fraction of input to set aside for evaluation set.
Returns:
A pipeline.Pipeline instance.
"""
# Transpose up to a major third in either direction.
transposition_range = list(range(-4, 5))
partitioner = pipelines_common.RandomPartition(
music_pb2.NoteSequence,
['eval_pianoroll_tracks', 'training_pianoroll_tracks'], [eval_ratio])
dag = {partitioner: dag_pipeline.DagInput(music_pb2.NoteSequence)}
for mode in ['eval', 'training']:
time_change_splitter = note_sequence_pipelines.TimeChangeSplitter(
name='TimeChangeSplitter_' + mode)
quantizer = note_sequence_pipelines.Quantizer(
steps_per_quarter=config.steps_per_quarter,
name='Quantizer_' + mode)
transposition_pipeline = note_sequence_pipelines.TranspositionPipeline(
transposition_range, name='TranspositionPipeline_' + mode)
pianoroll_extractor = PianorollSequenceExtractor(
min_steps=min_steps,
max_steps=max_steps,
name='PianorollExtractor_' + mode)
encoder_pipeline = event_sequence_pipeline.EncoderPipeline(
mm.PianorollSequence,
config.encoder_decoder,
name='EncoderPipeline_' + mode)
dag[time_change_splitter] = partitioner[mode + '_pianoroll_tracks']
dag[quantizer] = time_change_splitter
dag[transposition_pipeline] = quantizer
dag[pianoroll_extractor] = transposition_pipeline
dag[encoder_pipeline] = pianoroll_extractor
dag[dag_pipeline.DagOutput(mode +
'_pianoroll_tracks')] = encoder_pipeline
return dag_pipeline.DAGPipeline(dag)
def testQuantizer(self):
steps_per_quarter = 4
note_sequence = common_testing_lib.parse_test_proto(
music_pb2.NoteSequence,
"""
time_signatures: {
numerator: 4
denominator: 4}
tempos: {
qpm: 60}""")
testing_lib.add_track_to_sequence(
note_sequence, 0,
[(12, 100, 0.01, 10.0), (11, 55, 0.22, 0.50), (40, 45, 2.50, 3.50),
(55, 120, 4.0, 4.01), (52, 99, 4.75, 5.0)])
expected_quantized_sequence = sequences_lib.quantize_note_sequence(
note_sequence, steps_per_quarter)
unit = note_sequence_pipelines.Quantizer(steps_per_quarter)
self._unit_transform_test(unit, note_sequence,
[expected_quantized_sequence])
