def main():
# parse arguments
parser = argparse.ArgumentParser()
parser.add_argument('--timesteps',
help="model's range (default: %(default)s)",
type=int,
default=16)
parser.add_argument(
'-b',
'--batch_size_train',
help='batch size used during training phase (default: %(default)s)',
type=int,
default=128)
parser.add_argument(
'-s',
'--samples_per_epoch',
help='number of samples per epoch (default: %(default)s)',
type=int,
default=12800 * 7)
parser.add_argument(
'--num_val_samples',
help='number of validation samples (default: %(default)s)',
type=int,
default=1280)
parser.add_argument('-u',
'--num_units_lstm',
nargs='+',
help='number of lstm units (default: %(default)s)',
type=int,
default=[200, 200])
parser.add_argument(
'-d',
'--num_dense',
help='size of non recurrent hidden layers (default: %(default)s)',
type=int,
default=200)
parser.add_argument('-n',
'--name',
help='model name (default: %(default)s)',
choices=['deepbach', 'skip', 'norelu'],
type=str,
default='skip')
parser.add_argument(
'-i',
'--num_iterations',
help='number of gibbs iterations (default: %(default)s)',
type=int,
default=20000)
parser.add_argument('-t',
'--train',
nargs='?',
help='train models for N epochs (default: 15)',
default=0,
const=15,
type=int)
parser.add_argument('-p',
'--parallel',
nargs='?',
help='number of parallel updates (default: 16)',
type=int,
const=16,
default=1)
parser.add_argument('--overwrite',
help='overwrite previously computed models',
action='store_true')
parser.add_argument('-m',
'--midi_file',
nargs='?',
help='relative path to midi file',
type=str,
const='datasets/god_save_the_queen.mid')
parser.add_argument('-l',
'--length',
help='length of unconstrained generation',
type=int,
default=160)
parser.add_argument('--ext',
help='extension of model name',
type=str,
default='')
parser.add_argument('-o',
'--output_file',
nargs='?',
help='path to output file',
type=str,
default='',
const='generated_examples/example.mid')
parser.add_argument('--dataset',
nargs='?',
help='path to dataset folder',
type=str,
default='')
parser.add_argument(
'-r',
'--reharmonization',
nargs='?',
help='reharmonization of a melody from the corpus identified by its id',
type=int)
args = parser.parse_args()
print(args)
if args.ext:
ext = '_' + args.ext
else:
ext = ''
dataset_path = None
pickled_dataset = BACH_DATASET
# metadatas = [TickMetadatas(SUBDIVISION), FermataMetadatas(), KeyMetadatas(window_size=1)]
metadatas = [TickMetadatas(SUBDIVISION), FermataMetadatas()]
timesteps = args.timesteps
batch_size = args.batch_size_train
samples_per_epoch = args.samples_per_epoch
nb_val_samples = args.num_val_samples
num_units_lstm = args.num_units_lstm
model_name = args.name.lower() + ext
sequence_length = args.length
batch_size_per_voice = args.parallel
num_units_lstm = args.num_units_lstm
num_dense = args.num_dense
if args.output_file:
output_file = args.output_file
else:
output_file = None
parallel = batch_size_per_voice > 1
train = args.train > 0
num_epochs = args.train
overwrite = args.overwrite
# Create pickled dataset
if not os.path.exists(pickled_dataset):
initialization(dataset_path,
metadatas=metadatas,
voice_ids=[SOP_INDEX],
BACH_DATASET=BACH_DATASET)
# load dataset
X, X_metadatas, voice_ids, index2notes, note2indexes, metadatas = pickle.load(
open(pickled_dataset, 'rb'))
# dataset dependant variables
NUM_VOICES = len(voice_ids)
num_voices = NUM_VOICES
num_pitches = list(map(len, index2notes))
num_iterations = args.num_iterations // batch_size_per_voice // num_voices
# Create, train load models
if not os.path.exists('models/' + model_name + '_' + str(NUM_VOICES - 1) +
'.yaml'):
create_models(model_name,
create_new=overwrite,
num_units_lstm=num_units_lstm,
num_dense=num_dense,
pickled_dataset=pickled_dataset,
num_voices=num_voices,
metadatas=metadatas,
timesteps=timesteps)
if train:
models = train_models(model_name=model_name,
samples_per_epoch=samples_per_epoch,
num_epochs=num_epochs,
nb_val_samples=nb_val_samples,
timesteps=timesteps,
pickled_dataset=pickled_dataset,
num_voices=NUM_VOICES,
metadatas=metadatas,
batch_size=batch_size)
else:
models = load_models(model_name, num_voices=NUM_VOICES)
# todo to remove
# model_name = 'skip_large'
# timesteps = 32
#
# test_autoencoder(model_name='models/' + model_name + '_0',
# timesteps=timesteps,
# pickled_dataset=pickled_dataset)
distance_model = load_model('models/seq2seq_masking')
# distance_model.compile(optimizer='adam', loss='categorical_crossentropy',
# metrics=['accuracy'])
hidden_repr_model = Model(input=distance_model.input,
output=distance_model.layers[1].output)
hidden_repr_model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
# create target
left_features, _, _, _ = all_features(np.transpose(X[21], axes=(1, 0)),
voice_index=0,
time_index=16 * 4,
timesteps=32,
num_pitches=num_pitches,
num_voices=num_voices)
left_metas, central_metas, _ = all_metadatas(X_metadatas[21],
time_index=16 * 4,
timesteps=32,
metadatas=metadatas)
inputs_target_chorale = {
'left_features': np.array([left_features]),
'left_metas': np.array([left_metas]),
'central_metas': np.array([central_metas])
}
# show target
score = indexed_chorale_to_score(X[21][:, 16 * 4 - 32:16 * 4],
pickled_dataset=pickled_dataset)
score.show()
generated_chorale = gibbs(generation_models=models,
hidden_repr_model=hidden_repr_model,
inputs_target_chorale=inputs_target_chorale,
chorale_metas=X_metadatas[12][:150],
num_iterations=200,
pickled_dataset=pickled_dataset,
timesteps=timesteps)
# convert
score = indexed_chorale_to_score(np.transpose(generated_chorale,
axes=(1, 0)),
pickled_dataset=pickled_dataset)
score.show()
seqs = []
for k, chorale in enumerate(chorale_list):
seqs.append(chorale[:, :, :F_INDEX])
score = seqs_to_stream(seqs)
mf = midi.translate.music21ObjectToMidiFile(score)
midi_file_name = export_directory + 'all_chorales' + ".mid"
mf.open(midi_file_name, 'wb')
mf.write()
mf.close()
if __name__ == '__main__':
if not os.path.exists(RAW_DATASET):
initialization()
# load dataset
X, min_pitches, max_pitches, num_voices = pickle.load(
open(RAW_DATASET, 'rb'))
# parse arguments
parser = argparse.ArgumentParser()
parser.add_argument('--timesteps',
help="model's range (default: %(default)s)",
type=int,
default=16)
parser.add_argument(
'-b',
'--batch_size_train',
help='batch size used during training phase (default: %(default)s)',
def main():
# parse arguments
parser = argparse.ArgumentParser()
parser.add_argument('--timesteps',
help="model's range (default: %(default)s)",
type=int,
default=16)
parser.add_argument(
'-b',
'--batch_size_train',
help='batch size used during training phase (default: %(default)s)',
type=int,
default=128)
parser.add_argument(
'-s',
'--steps_per_epoch',
help='number of steps per epoch (default: %(default)s)',
type=int,
default=500)
parser.add_argument(
'--validation_steps',
help='number of validation steps (default: %(default)s)',
type=int,
default=20)
parser.add_argument('-u',
'--num_units_lstm',
nargs='+',
help='number of lstm units (default: %(default)s)',
type=int,
default=[200, 200])
parser.add_argument(
'-d',
'--num_dense',
help='size of non recurrent hidden layers (default: %(default)s)',
type=int,
default=200)
parser.add_argument('-n',
'--name',
help='model name (default: %(default)s)',
choices=['deepbach', 'skip'],
type=str,
default='deepbach')
parser.add_argument(
'-i',
'--num_iterations',
help='number of gibbs iterations (default: %(default)s)',
type=int,
default=20000)
parser.add_argument('-t',
'--train',
nargs='?',
help='train models for N epochs (default: 15)',
default=0,
const=15,
type=int)
parser.add_argument('-p',
'--parallel',
nargs='?',
help='number of parallel updates (default: 16)',
type=int,
const=16,
default=1)
parser.add_argument('--overwrite',
help='overwrite previously computed models',
action='store_true')
parser.add_argument('-m',
'--midi_file',
nargs='?',
help='relative path to midi file',
type=str,
const='datasets/god_save_the_queen.mid')
parser.add_argument('-l',
'--length',
help='length of unconstrained generation',
type=int,
default=160)
parser.add_argument('--ext',
help='extension of model name',
type=str,
default='')
parser.add_argument('-o',
'--output_file',
nargs='?',
help='path to output file',
type=str,
default='',
const='generated_examples/example.mid')
parser.add_argument('--dataset',
nargs='?',
help='path to dataset folder',
type=str,
default='')
parser.add_argument(
'-r',
'--reharmonization',
nargs='?',
help='reharmonization of a melody from the corpus identified by its id',
type=int)
args = parser.parse_args()
print(args)
# fixed set of metadatas to use when CREATING the dataset
# Available metadatas:
# metadatas = [FermataMetadatas(), KeyMetadatas(window_size=1), TickMetadatas(SUBDIVISION), ModeMetadatas()]
metadatas = [
TickMetadatas(SUBDIVISION),
FermataMetadatas(),
KeyMetadatas(window_size=1)
]
if args.ext:
ext = '_' + args.ext
else:
ext = ''
# datasets
# set pickled_dataset argument
if args.dataset:
dataset_path = args.dataset
dataset_name = dataset_path.split('/')[-1]
pickled_dataset = 'datasets/custom_dataset/' + dataset_name + '.pickle'
else:
dataset_path = None
pickled_dataset = BACH_DATASET
if not os.path.exists(pickled_dataset):
initialization(dataset_path,
metadatas=metadatas,
voice_ids=[0, 1, 2, 3])
# load dataset
X, X_metadatas, voice_ids, index2notes, note2indexes, metadatas = pickle.load(
open(pickled_dataset, 'rb'))
NUM_VOICES = len(voice_ids)
num_pitches = list(map(len, index2notes))
timesteps = args.timesteps
batch_size = args.batch_size_train
steps_per_epoch = args.steps_per_epoch
validation_steps = args.validation_steps
num_units_lstm = args.num_units_lstm
model_name = args.name.lower() + ext
sequence_length = args.length
batch_size_per_voice = args.parallel
num_units_lstm = args.num_units_lstm
num_dense = args.num_dense
if args.output_file:
output_file = args.output_file
else:
output_file = None
# when reharmonization
if args.midi_file:
melody = converter.parse(args.midi_file)
melody = part_to_inputs(melody.parts[0],
index2note=index2notes[0],
note2index=note2indexes[0])
num_voices = NUM_VOICES - 1
# todo find a way to specify metadatas when reharmonizing a given melody
chorale_metas = [
metas.generate(sequence_length) for metas in metadatas
]
elif args.reharmonization:
melody = X[args.reharmonization][0, :]
num_voices = NUM_VOICES - 1
chorale_metas = X_metadatas[args.reharmonization]
else:
num_voices = NUM_VOICES
melody = None
# todo find a better way to set metadatas
# chorale_metas = [metas[:sequence_length] for metas in X_metadatas[11]]
chorale_metas = [
metas.generate(sequence_length) for metas in metadatas
]
num_iterations = args.num_iterations // batch_size_per_voice // num_voices
parallel = batch_size_per_voice > 1
train = args.train > 0
num_epochs = args.train
overwrite = args.overwrite
if not os.path.exists('models/' + model_name + '_' + str(NUM_VOICES - 1) +
'.yaml'):
create_models(model_name,
create_new=overwrite,
num_units_lstm=num_units_lstm,
num_dense=num_dense,
pickled_dataset=pickled_dataset,
num_voices=num_voices,
metadatas=metadatas,
timesteps=timesteps)
if train:
models = train_models(model_name=model_name,
steps_per_epoch=steps_per_epoch,
num_epochs=num_epochs,
validation_steps=validation_steps,
timesteps=timesteps,
pickled_dataset=pickled_dataset,
num_voices=NUM_VOICES,
metadatas=metadatas,
batch_size=batch_size)
else:
models = load_models(model_name, num_voices=NUM_VOICES)
temperature = 1.
timesteps = int(models[0].input[0]._keras_shape[1])
seq = generation(model_base_name=model_name,
models=models,
timesteps=timesteps,
melody=melody,
initial_seq=None,
temperature=temperature,
chorale_metas=chorale_metas,
parallel=parallel,
batch_size_per_voice=batch_size_per_voice,
num_iterations=num_iterations,
sequence_length=sequence_length,
output_file=output_file,
pickled_dataset=pickled_dataset)
def main():
# parse arguments
parser = argparse.ArgumentParser()
parser.add_argument('--timesteps',
help="model's range (default: %(default)s)",
type=int,
default=16)
parser.add_argument(
'-b',
'--batch_size_train',
help='batch size used during training phase (default: %(default)s)',
type=int,
default=128)
parser.add_argument(
'-s',
'--steps_per_epoch',
help='number of steps per epoch (default: %(default)s)',
type=int,
default=500)
parser.add_argument(
'--validation_steps',
help='number of validation steps (default: %(default)s)',
type=int,
default=20)
parser.add_argument('-u',
'--num_units_lstm',
nargs='+',
help='number of lstm units (default: %(default)s)',
type=int,
default=[200, 200])
parser.add_argument(
'-d',
'--num_dense',
help='size of non recurrent hidden layers (default: %(default)s)',
type=int,
default=200)
parser.add_argument('-n',
'--name',
help='model name (default: %(default)s)',
choices=['deepbach', 'skip'],
type=str,
default='deepbach')
parser.add_argument(
'-i',
'--num_iterations',
help='number of gibbs iterations (default: %(default)s)',
type=int,
default=6000)
parser.add_argument('-t',
'--train',
nargs='?',
help='train models for N epochs (default: 15)',
default=0,
const=15,
type=int)
parser.add_argument('-p',
'--parallel',
nargs='?',
help='number of parallel updates (default: 16)',
type=int,
const=16,
default=1)
parser.add_argument('--overwrite',
help='overwrite previously computed models',
action='store_true')
parser.add_argument('-m',
'--midi_file',
nargs='?',
help='relative path to midi file',
type=str,
const='datasets/god_save_the_queen.mid')
parser.add_argument('-l',
'--length',
help='length of unconstrained generation',
type=int,
default=256)
parser.add_argument('--ext',
help='extension of model name',
type=str,
default='')
parser.add_argument('-o',
'--output_file',
nargs='?',
help='path to output file',
type=str,
default='',
const='generated_examples/example.mid')
parser.add_argument(
'-r',
'--reharmonization',
nargs='?',
help='reharmonization of a melody from the corpus identified by its id',
type=int)
args = parser.parse_args()
print(args)
# fixed set of metadatas to use when CREATING the dataset
# Available metadatas:
# metadatas = [FermataMetadatas(), KeyMetadatas(window_size=1), TickMetadatas(SUBDIVISION), ModeMetadatas()]
metadatas = [TickMetadatas(SUBDIVISION)
] #, FermataMetadatas(), KeyMetadatas(window_size=1)]
if args.ext:
ext = '_' + args.ext
else:
ext = ''
# datasets
# set pickled_dataset argument
#dataset_path = None
pickled_dataset = ARASHI_DATASET
if not os.path.exists(pickled_dataset):
initialization(metadatas=metadatas, voice_ids=[0, 1])
# load dataset
X, X_metadatas, voice_ids, index2notes, note2indexes, metadatas, chordname2chordset = pickle.load(
open(pickled_dataset, 'rb'))
#index2notes：サイズがパート数の配列、各要素はset型、setの中身は 各パートに出てきた音と番号の対応表
print('index2notes' + '=' * 50)
print(index2notes)
#print('X'+'='*50)
#print(X[0][1])
#print('X_metadatas'+'='*50)
#chorale = music21.corpus.parse('bach/bwv86.6')
#chorale.show()
#length = int(chorale.duration.quarterLength * SUBDIVISION)
#print(length)
#print(np.array(list(map(lambda x: x % 5,range(length)))))
#print(type(X_metadatas))
#TickMetadatas(X_metadatas).evaluate(chorale)
NUM_VOICES = len(voice_ids)
num_pitches = list(map(
len, index2notes)) #パートごとの全曲の音の数(+END,STARTとか)、サイズは(パート数)
timesteps = args.timesteps
batch_size = args.batch_size_train
steps_per_epoch = args.steps_per_epoch
validation_steps = args.validation_steps
num_units_lstm = args.num_units_lstm
model_name = args.name.lower() + ext
sequence_length = args.length
batch_size_per_voice = args.parallel
num_units_lstm = args.num_units_lstm
num_dense = args.num_dense
#print('num_pitches = ', list(map(len, index2notes)))
#print('NUM_VOICES = ', len(voice_ids))
if args.output_file:
output_file = args.output_file
else:
output_file = None
# when reharmonization
if args.midi_file:
melody = converter.parse(args.midi_file)
melody = melody_to_inputs(melody.parts[0],
int(melody.parts[0].duration.quarterLength *
SUBDIVISION),
index2note=index2notes[0],
note2index=note2indexes[0])
num_voices = NUM_VOICES - 1
sequence_length = len(melody)
# todo find a way to specify metadatas when reharmonizing a given melody
chorale_metas = [
metas.generate(sequence_length) for metas in metadatas
]
elif args.reharmonization:
melody = X[args.reharmonization][0, :]
num_voices = NUM_VOICES - 1
chorale_metas = X_metadatas[args.reharmonization]
else:
num_voices = NUM_VOICES
melody = None
# todo find a better way to set metadatas
# chorale_metas = [metas[:sequence_length] for metas in X_metadatas[11]]
chorale_metas = [
metas.generate(sequence_length) for metas in metadatas
] #[TickMetadatas(SUBDIVISION), FermataMetadatas(), KeyMetadatas(window_size=1)]の各要素について.generateを実行、長さ=length(=160)のarray×3の配列を作る
#print('sequence_length=', sequence_length )
#print('choral_metas=', chorale_metas)
num_iterations = args.num_iterations // batch_size_per_voice // num_voices #iteration回数を各パートで分配
print('args.num_iterations=', args.num_iterations)
print('batch_size_per_voice=', batch_size_per_voice)
print('num_voices=', num_voices)
parallel = batch_size_per_voice > 1
print(' parallel=', parallel)
train = args.train > 0
num_epochs = args.train #-t で指定した値
print(' num_epochs=', num_epochs)
overwrite = args.overwrite
#print(timesteps)
if not os.path.exists('models/' + model_name + '_' + str(NUM_VOICES - 1) +
'.yaml'):
create_models(model_name,
create_new=overwrite,
num_units_lstm=num_units_lstm,
num_dense=num_dense,
pickled_dataset=pickled_dataset,
num_voices=num_voices,
metadatas=metadatas,
timesteps=timesteps)
if train:
models = train_models(model_name=model_name,
steps_per_epoch=steps_per_epoch,
num_epochs=num_epochs,
validation_steps=validation_steps,
timesteps=timesteps,
pickled_dataset=pickled_dataset,
num_voices=NUM_VOICES,
metadatas=metadatas,
batch_size=batch_size)
else:
models = load_models(model_name, num_voices=NUM_VOICES)
temperature = 1.
timesteps = int(models[0].input[0]._keras_shape[1])
print('==============generation==============')
seq = generation(model_base_name=model_name,
models=models,
timesteps=timesteps,
melody=melody,
initial_seq=None,
temperature=temperature,
chorale_metas=chorale_metas,
parallel=parallel,
batch_size_per_voice=batch_size_per_voice,
num_iterations=num_iterations,
sequence_length=sequence_length,
output_file=output_file,
pickled_dataset=pickled_dataset)