def prime_model(self):
"""Primes the model with its default midi primer."""
with self.graph.as_default():
tf.logging.debug('Priming the model with MIDI file %s',
self.midi_primer)
# Convert primer Melody to model inputs.
encoder = note_seq.OneHotEventSequenceEncoderDecoder(
note_seq.MelodyOneHotEncoding(min_note=rl_tuner_ops.MIN_NOTE,
max_note=rl_tuner_ops.MAX_NOTE))
primer_input, _ = encoder.encode(self.primer)
# Run model over primer sequence.
primer_input_batch = np.tile([primer_input],
(self.batch_size, 1, 1))
self.state_value, softmax = self.session.run(
[self.state_tensor, self.softmax],
feed_dict={
self.initial_state:
self.state_value,
self.melody_sequence:
primer_input_batch,
self.lengths:
np.full(self.batch_size, len(self.primer), dtype=int)
})
priming_output = softmax[-1, :]
self.priming_note = self.get_note_from_softmax(priming_output)
def convert_to_note_events(one_hot_events):
_, result = np.where(one_hot_events == 1.)
encoder_decoder = note_seq.MelodyOneHotEncoding(BASIC_DEFAULT_MIN_NOTE,
BASIC_DEFAULT_MAX_NOTE)
result = np.array(
[encoder_decoder.decode_event(event) for event in result])
return result
def setUp(self):
super().setUp()
self.config = melody_rnn_model.MelodyRnnConfig(
None,
note_seq.OneHotEventSequenceEncoderDecoder(
note_seq.MelodyOneHotEncoding(0, 127)),
contrib_training.HParams(),
min_note=0,
max_note=127,
transpose_to_key=0)
def lookback_melody_encoder_decoder(min_note, max_note):
"""Return a LookbackEventSequenceEncoderDecoder for melodies.
Args:
min_note: The minimum midi pitch the encoded melodies can have.
max_note: The maximum midi pitch (exclusive) the encoded melodies can have.
Returns:
A melody LookbackEventSequenceEncoderDecoder.
"""
return note_seq.LookbackEventSequenceEncoderDecoder(
note_seq.MelodyOneHotEncoding(min_note, max_note))
def setUp(self):
super().setUp()
self.config = improv_rnn_model.ImprovRnnConfig(
None,
note_seq.ConditionalEventSequenceEncoderDecoder(
note_seq.OneHotEventSequenceEncoderDecoder(
note_seq.MajorMinorChordOneHotEncoding()),
note_seq.OneHotEventSequenceEncoderDecoder(
note_seq.MelodyOneHotEncoding(0, 127))),
contrib_training.HParams(),
min_note=0,
max_note=127,
transpose_to_key=0)
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 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)])
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 = note_seq.OneHotEventSequenceEncoderDecoder(
note_seq.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 = pipelines_common.make_sequence_example(
*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 __init__(self, min_pitch, max_pitch):
self._encoding = note_seq.MelodyOneHotEncoding(min_note=min_pitch,
max_note=max_pitch + 1)
self.transpose_to_key = transpose_to_key
# Default configurations.
default_configs = {
'basic_improv':
ImprovRnnConfig(
generator_pb2.GeneratorDetails(
id='basic_improv',
description='Basic melody-given-chords RNN with one-hot triad '
'encoding for chords.'),
note_seq.ConditionalEventSequenceEncoderDecoder(
note_seq.OneHotEventSequenceEncoderDecoder(
note_seq.TriadChordOneHotEncoding()),
note_seq.OneHotEventSequenceEncoderDecoder(
note_seq.MelodyOneHotEncoding(min_note=DEFAULT_MIN_NOTE,
max_note=DEFAULT_MAX_NOTE))),
contrib_training.HParams(batch_size=128,
rnn_layer_sizes=[64, 64],
dropout_keep_prob=0.5,
clip_norm=5,
learning_rate=0.001)),
'attention_improv':
ImprovRnnConfig(
generator_pb2.GeneratorDetails(
id='attention_improv',
description=
'Melody-given-chords RNN with one-hot triad encoding for '
'chords, attention, and binary counters.'),
note_seq.ConditionalEventSequenceEncoderDecoder(
note_seq.OneHotEventSequenceEncoderDecoder(
note_seq.TriadChordOneHotEncoding()),
