def get_config(batch_size, data_path):
return configs.Config(
model=MusicVAE(lstm_models.BidirectionalLstmEncoder(),
lstm_models.CategoricalLstmDecoder()),
hparams=merge_hparams(
lstm_models.get_default_hparams(),
HParams(
batch_size=512,
max_seq_len=32, # 2 bars w/ 16 steps per bar
z_size=512,
enc_rnn_size=[2048],
dec_rnn_size=[2048, 2048, 2048],
free_bits=0,
max_beta=0.5,
beta_rate=0.99999,
sampling_schedule='inverse_sigmoid',
sampling_rate=1000,
)),
note_sequence_augmenter=data.NoteSequenceAugmenter(
transpose_range=(-5, 5)),
data_converter=data.OneHotMelodyConverter(
valid_programs=data.MEL_PROGRAMS,
skip_polyphony=False,
max_bars=100, # Truncate long melodies before slicing.
slice_bars=2,
steps_per_quarter=4),
train_examples_path=data_path,
eval_examples_path=data_path,
)
def testTfAugment(self):
augmenter = data.NoteSequenceAugmenter(transpose_range=(-3, -3),
stretch_range=(2.0, 2.0))
augmented_sequence = augmenter.augment(self.sequence)
with self.test_session() as sess:
sequence_str = tf.placeholder(tf.string)
augmented_sequence_str_ = augmenter.tf_augment(sequence_str)
augmented_sequence_str = sess.run(
[augmented_sequence_str_],
feed_dict={sequence_str: self.sequence.SerializeToString()})
augmented_sequence = music_pb2.NoteSequence.FromString(
augmented_sequence_str[0])
expected_sequence = music_pb2.NoteSequence()
expected_sequence.tempos.add(qpm=30)
testing_lib.add_track_to_sequence(expected_sequence, 0,
[(29, 100, 4, 8), (30, 100, 12, 22),
(31, 100, 22, 26),
(32, 100, 34, 36)])
testing_lib.add_track_to_sequence(expected_sequence,
1, [(57, 80, 8, 8.2),
(58, 80, 24, 24.2)],
is_drum=True)
testing_lib.add_chords_to_sequence(expected_sequence, [('N.C.', 0),
('A', 16),
('Gbm', 32)])
self.assertEqual(expected_sequence, augmented_sequence)
def testAugmentStretch(self):
augmenter = data.NoteSequenceAugmenter(stretch_range=(0.5, 0.5))
augmented_sequence = augmenter.augment(self.sequence)
expected_sequence = music_pb2.NoteSequence()
expected_sequence.tempos.add(qpm=120)
testing_lib.add_track_to_sequence(
expected_sequence, 0,
[(32, 100, 1, 2), (33, 100, 3, 5.5), (34, 100, 5.5, 6.5),
(35, 100, 8.5, 9)])
testing_lib.add_track_to_sequence(
expected_sequence, 1, [(57, 80, 2, 2.05), (58, 80, 6, 6.05)],
is_drum=True)
testing_lib.add_chords_to_sequence(
expected_sequence, [('N.C.', 0), ('C', 4), ('Am', 8)])
self.assertEqual(expected_sequence, augmented_sequence)
def testAugmentTranspose(self):
augmenter = data.NoteSequenceAugmenter(transpose_range=(2, 2))
augmented_sequence = augmenter.augment(self.sequence)
expected_sequence = music_pb2.NoteSequence()
expected_sequence.tempos.add(qpm=60)
testing_lib.add_track_to_sequence(
expected_sequence, 0,
[(34, 100, 2, 4), (35, 100, 6, 11), (36, 100, 11, 13),
(37, 100, 17, 18)])
testing_lib.add_track_to_sequence(
expected_sequence, 1, [(57, 80, 4, 4.1), (58, 80, 12, 12.1)],
is_drum=True)
testing_lib.add_chords_to_sequence(
expected_sequence, [('N.C.', 0), ('D', 8), ('Bm', 16)])
self.assertEqual(expected_sequence, augmented_sequence)
lstm_models.CategoricalLstmDecoder()),
hparams=merge_hparams(
lstm_models.get_default_hparams(),
HParams(
batch_size=512,
max_seq_len=32, # 2 bars w/ 16 steps per bar
z_size=256,
enc_rnn_size=[512],
dec_rnn_size=[256, 256],
free_bits=0,
max_beta=0.2,
beta_rate=0.99999,
sampling_schedule='inverse_sigmoid',
sampling_rate=1000,
)),
note_sequence_augmenter=data.NoteSequenceAugmenter(transpose_range=(-5,
5)),
data_converter=data.OneHotMelodyConverter(
valid_programs=data.MEL_PROGRAMS,
skip_polyphony=False,
max_bars=100, # Truncate long melodies before slicing.
slice_bars=2,
steps_per_quarter=4),
train_examples_path=None,
eval_examples_path=None,
)
CONFIG_MAP['cat-mel_2bar_big'] = Config(
model=MusicVAE(lstm_models.BidirectionalLstmEncoder(),
lstm_models.CategoricalLstmDecoder()),
hparams=merge_hparams(
lstm_models.get_default_hparams(),
