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,
)
Config.__new__.__defaults__ = (None, ) * len(Config._fields)
def update_config(config, update_dict):
config_dict = config.values()
config_dict.update(update_dict)
return Config(**config_dict)
CONFIG_MAP = {}
# Melody
CONFIG_MAP['cat-mel_2bar_small'] = 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=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,
])):
def values(self):
return self._asdict()
def update_config(config, update_dict):
config_dict = config.values()
config_dict.update(update_dict)
return Config(**config_dict)
config_map = {}
# Melody
config_map['cat-mel_2bar_small'] = 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=256,
enc_rnn_size=[512],
dec_rnn_size=[256, 256],
)),
note_sequence_augmenter=None,
note_sequence_converter=data.OneHotMelodyConverter(
valid_programs=data.MEL_PROGRAMS,
skip_polyphony=True,
max_bars=100, # Truncate long melodies before slicing.
slice_bars=2,
note3_set.append(sequence)
# note_seq.play_sequence(sequence, synth=note_seq.synthesize)
count += 1
# initializes a 16 bar trio model to train
trio_16bar_converter = data.TrioConverter(steps_per_quarter=4,
slice_bars=16,
gap_bars=2)
# simplest version of the 16 bar trio because it has the largest size
# and I can understand it best
CONFIG_MAP['flat-trio_16bar'] = Config(
model=MusicVAE(
lstm_models.BidirectionalLstmEncoder(),
lstm_models.MultiOutCategoricalLstmDecoder(output_depths=[
90, # melody
90, # bass
512, # drums
])),
hparams=merge_hparams(
lstm_models.get_default_hparams(),
HParams(
batch_size=256,
max_seq_len=256,
z_size=512,
enc_rnn_size=[2048, 2048],
dec_rnn_size=[2048, 2048, 2048],
)),
note_sequence_augmenter=None,
data_converter=trio_16bar_converter,
train_examples_path='./MidiSet1/',