def create_models(model_name=None,
create_new=False,
num_dense=200,
num_units_lstm=[200, 200],
pickled_dataset=BACH_DATASET,
num_voices=4,
metadatas=None,
timesteps=16):
"""
Choose one model
:param model_name:
:return:
"""
_, _, _, index2notes, _, _ = pickle.load(open(pickled_dataset, 'rb'))
num_pitches = list(map(len, index2notes))
for voice_index in range(num_voices):
# We only need one example for features dimensions
gen = generator_from_raw_dataset(batch_size=1,
timesteps=timesteps,
voice_index=voice_index,
pickled_dataset=pickled_dataset)
((left_features, central_features, right_features),
(left_metas, central_metas, right_metas), labels) = next(gen)
if 'deepbach' in model_name:
model = deepBach(num_features_lr=left_features.shape[-1],
num_features_c=central_features.shape[-1],
num_pitches=num_pitches[voice_index],
num_features_meta=left_metas.shape[-1],
num_dense=num_dense,
num_units_lstm=num_units_lstm)
elif 'skip' in model_name:
model = deepbach_skip_connections(
num_features_lr=left_features.shape[-1],
num_features_c=central_features.shape[-1],
num_features_meta=left_metas.shape[-1],
num_pitches=num_pitches[voice_index],
num_dense=num_dense,
num_units_lstm=num_units_lstm,
timesteps=timesteps)
else:
raise ValueError
model_path_name = 'models/' + model_name + '_' + str(voice_index)
if not os.path.exists(model_path_name + '.json') or create_new:
save_model(model, model_name=model_path_name, overwrite=create_new)
def create_models(model_name=None,
create_new=False,
num_dense=200,
num_units_lstm=[200, 200]):
"""
Choose one model
:param model_name:
:return:
"""
_, min_pitches, max_pitches, num_voices = pickle.load(
open(RAW_DATASET, 'rb'))
for voice_index in range(4):
# We only need one example for features dimensions
gen = generator_from_raw_dataset(batch_size=batch_size,
timesteps=timesteps,
voice_index=voice_index)
(left_features, central_features, right_features, beats, labels,
fermatas) = next(gen)
if 'deepbach' in model_name:
model = deepBach(
num_features_lr=left_features.shape[-1],
num_features_c=central_features.shape[-1],
num_pitches=max_pitches[voice_index] -
min_pitches[voice_index] + 1 + 1, # for continuation symbol
num_dense=num_dense,
num_units_lstm=num_units_lstm)
elif 'maxent' in model_name:
model = maxEnt(num_features_lr=left_features.shape[-1],
num_features_c=central_features.shape[-1],
num_pitches=max_pitches[voice_index] -
min_pitches[voice_index] + 1 +
1) # for continuation symbol
elif 'mlp' in model_name:
model = mlp(num_features_lr=left_features.shape[-1],
num_features_c=central_features.shape[-1],
num_pitches=max_pitches[voice_index] -
min_pitches[voice_index] + 1 + 1,
num_hidden=num_dense) # for continuation symbol
else:
raise ValueError
model_path_name = 'models/' + model_name + '_' + str(voice_index)
if not os.path.exists(model_path_name + '.json') or create_new:
save_model(model, model_name=model_path_name, overwrite=create_new)
def test_autoencoder(model_name, timesteps, pickled_dataset=BACH_DATASET):
voice_index = 0
num_epochs = 200
samples_per_epoch = 1024 * 100
batch_size = 64
nb_val_samples = 1024
X, X_metadatas, voice_ids, index2notes, note2indexes, metadatas = pickle.load(
open(pickled_dataset, 'rb'))
# sequences
num_voices = 1
num_pitches = list(map(len, index2notes))
generator_train = (({
'left_features': left_features,
'central_features': central_features,
'right_features': right_features,
'left_metas': left_metas,
'right_metas': right_metas,
'central_metas': central_metas,
}, {
'pitch_prediction': labels
}) for (
(left_features, central_features,
right_features), (left_metas, central_metas, right_metas),
labels) in generator_from_raw_dataset(batch_size=batch_size,
timesteps=timesteps,
voice_index=voice_index,
phase='train',
pickled_dataset=pickled_dataset))
generator_unitary = (({
'left_features': left_features,
'central_features': central_features,
'right_features': right_features,
'left_metas': left_metas,
'right_metas': right_metas,
'central_metas': central_metas,
}, {
'pitch_prediction': labels
}) for (
(left_features, central_features,
right_features), (left_metas, central_metas, right_metas),
labels) in generator_from_raw_dataset(batch_size=1,
timesteps=timesteps,
voice_index=voice_index,
phase='all',
pickled_dataset=pickled_dataset))
inputs, outputs = next(generator_train)
model = load_model(model_name)
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
hidden_repr_model = Model(input=model.input,
output=model.layers[-1].output)
hidden_repr_model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
# create score target
chorale_seq = chorale_onehot_to_indexed_chorale(
onehot_chorale=inputs['left_features'][0],
num_pitches=num_pitches,
time_major=False)
score = indexed_chorale_to_score(chorale_seq,
pickled_dataset=pickled_dataset)
score.show()
nearest_chorale_inputs, intermediate_results = find_nearest(
inputs, hidden_repr_model, generator_unitary, num_elements=20000)
# concat all results
nearest_chorale = np.concatenate([
np.array(nearest_chorale_inputs[0]['left_features'][0])
for nearest_chorale_inputs in intermediate_results
],
axis=0)
# create score nearest
nearest_chorale_seq = chorale_onehot_to_indexed_chorale(
onehot_chorale=nearest_chorale,
num_pitches=num_pitches,
time_major=False)
score_nearest = indexed_chorale_to_score(nearest_chorale_seq,
pickled_dataset=pickled_dataset)
score_nearest.show()
def train_models(model_name,
steps_per_epoch,
num_epochs,
validation_steps,
timesteps,
pickled_dataset=BACH_DATASET,
num_voices=4,
batch_size=16,
metadatas=None):
"""
Train models
:param batch_size:
:param metadatas:
"""
models = []
for voice_index in range(num_voices):
# Load appropriate generators
generator_train = (({
'left_features': left_features,
'central_features': central_features,
'right_features': right_features,
'left_metas': left_metas,
'right_metas': right_metas,
'central_metas': central_metas,
}, {
'pitch_prediction': labels
}) for ((left_features, central_features,
right_features), (left_metas, central_metas, right_metas),
labels) in generator_from_raw_dataset(
batch_size=batch_size,
timesteps=timesteps,
voice_index=voice_index,
phase='train',
pickled_dataset=pickled_dataset))
generator_val = (({
'left_features': left_features,
'central_features': central_features,
'right_features': right_features,
'left_metas': left_metas,
'right_metas': right_metas,
'central_metas': central_metas,
}, {
'pitch_prediction': labels
}) for ((left_features, central_features,
right_features), (left_metas, central_metas, right_metas),
labels) in generator_from_raw_dataset(
batch_size=batch_size,
timesteps=timesteps,
voice_index=voice_index,
phase='test',
pickled_dataset=pickled_dataset))
model_path_name = 'models/' + model_name + '_' + str(voice_index)
model = load_model(model_path_name)
model.compile(optimizer='adam',
loss={'pitch_prediction': 'categorical_crossentropy'},
metrics=['accuracy'])
model.fit_generator(generator_train,
samples_per_epoch=steps_per_epoch,
epochs=num_epochs,
verbose=1,
validation_data=generator_val,
validation_steps=validation_steps)
models.append(model)
save_model(model, model_path_name, overwrite=True)
return models
def train_models(model_name, samples_per_epoch, num_epochs, nb_val_samples):
"""
Train models
"""
models = []
for voice_index in range(4):
# Load appropriate generators
generator_train = (({
'left_features': left_features,
'central_features': central_features,
'right_features': right_features,
'beat': beat,
'beats_left': beats_left,
'beats_right': beats_right,
'fermatas_left': fermatas_left,
'fermatas_right': fermatas_right,
'central_fermata': central_fermata
}, {
'pitch_prediction': labels
}) for (left_features, central_features, right_features,
(beats_left, beat, beats_right), labels,
(fermatas_left, central_fermata, fermatas_right
)) in generator_from_raw_dataset(batch_size=batch_size,
timesteps=timesteps,
voice_index=voice_index,
phase='train'))
generator_val = (({
'left_features': left_features,
'central_features': central_features,
'right_features': right_features,
'beat': beat,
'beats_left': beats_left,
'beats_right': beats_right,
'fermatas_left': fermatas_left,
'fermatas_right': fermatas_right,
'central_fermata': central_fermata
}, {
'pitch_prediction': labels
}) for (left_features, central_features, right_features,
(beats_left, beat, beats_right), labels,
(fermatas_left, central_fermata, fermatas_right
)) in generator_from_raw_dataset(batch_size=batch_size,
timesteps=timesteps,
voice_index=voice_index,
phase='train'))
model_path_name = 'models/' + model_name + '_' + str(voice_index)
model = load_model(model_path_name)
model.compile(optimizer='adam',
loss={'pitch_prediction': 'categorical_crossentropy'},
metrics=['accuracy'])
model.fit_generator(generator_train,
samples_per_epoch=samples_per_epoch,
nb_epoch=num_epochs,
verbose=1,
validation_data=generator_val,
nb_val_samples=nb_val_samples)
models.append(model)
save_model(model, model_path_name, overwrite=True)
return models
def train_models(model_name,
steps_per_epoch,
num_epochs,
validation_steps,
timesteps,
pickled_dataset=ARASHI_DATASET,
num_voices=4,
batch_size=16,
metadatas=None):
"""
Train models
:param batch_size:
:param metadatas:
"""
models = []
for voice_index in range(num_voices):
# Load appropriate generators
#print('髪切ったー'*100)
generator_train = (({
'left_features': left_features,
'central_features': central_features,
'right_features': right_features,
'left_metas': left_metas,
'right_metas': right_metas,
'central_metas': central_metas,
'left_local_seqs': left_local_seqs,
'right_local_seqs': right_local_seqs
}, {
'pitch_prediction': labels
}) for ((left_features, central_features,
right_features), (left_metas, central_metas,
right_metas), labels, left_local_seqs,
right_local_seqs) in generator_from_raw_dataset(
batch_size=batch_size,
timesteps=timesteps,
voice_index=voice_index,
phase='train',
pickled_dataset=pickled_dataset)) ########訓練データ
#print('春服買いたい'*100)
generator_val = (({
'left_features': left_features,
'central_features': central_features,
'right_features': right_features,
'left_metas': left_metas,
'right_metas': right_metas,
'central_metas': central_metas,
'left_local_seqs': left_local_seqs,
'right_local_seqs': right_local_seqs
}, {
'pitch_prediction': labels
}) for ((left_features, central_features,
right_features), (left_metas, central_metas,
right_metas), labels, left_local_seqs,
right_local_seqs) in generator_from_raw_dataset(
batch_size=batch_size,
timesteps=timesteps,
voice_index=voice_index,
phase='test',
pickled_dataset=pickled_dataset)) #######検証データ
#print('いちご食べたい'*100)
#print(model_name)
model_path_name = 'models/' + model_name + '_' + str(voice_index)
model = load_model(model_path_name)
model.compile(optimizer='adam',
loss={'pitch_prediction': 'categorical_crossentropy'},
metrics=['accuracy']) #############すべてを理解した(*ﾉω・*)ﾃﾍ
#print(validation_steps)
model.fit_generator(generator_train,
samples_per_epoch=steps_per_epoch,
epochs=num_epochs,
verbose=1,
validation_data=generator_val,
validation_steps=validation_steps)
models.append(model) #modelという要素をもつリストを作った
save_model(model, model_path_name, overwrite=True)
#print('はい'*100)
#print(len(models))
return models
def create_models(model_name=None,
create_new=False,
num_dense=200,
num_units_lstm=[200, 200],
pickled_dataset=ARASHI_DATASET,
num_voices=4,
metadatas=None,
timesteps=16):
"""
Choose one model
:param model_name:
:return:
"""
_, _, _, index2notes, _, _, _ = pickle.load(open(pickled_dataset, 'rb'))
# X, X_metadatas, voice_ids, index2notes, note2indexes, metadata = pickle.load(open(pickled_dataset, 'rb'))
#print('統計物理重い'*50)
#print('index2notes===================\n', index2notes)
num_pitches = list(map(len, index2notes)) #各パートの音の種類数
#print('num_pitches=',num_pitches)
for voice_index in range(num_voices):
print('=' * 10 + 'パート', voice_index, '=' * 10, timesteps)
# We only need one example for features dimensions
gen = generator_from_raw_dataset(batch_size=1,
timesteps=timesteps,
voice_index=voice_index,
pickled_dataset=pickled_dataset)
#batch_size=1としているのはgenerator_from_raw_datasetのwhile文の反復を１回にするため、ここではモデルのサイズさえわかれば良いので
((left_features, central_features,
right_features), (left_metas, central_metas, right_metas), labels,
left_local_seqs, right_local_seqs) = next(gen)
"""
print()
print('num_features_lr=', left_features.shape[-1])
print('num_features_c=', central_features.shape[-1])
print('num_pitches=', num_pitches)
print('num_features_meta=', left_metas.shape[-1])
print('num_dense=', num_dense)
print('num_units_lstm=', num_units_lstm)
print()
"""
#if voice_index == 0:
#np.set_printoptions(threshold=np.inf)
#print('bbbbbbbbbb',left_features.shape)
#print(left_features)
if 'deepbach' in model_name:
#model = deepBach(num_features_lr=left_features.shape[-1],
# num_features_c=central_features.shape[-1],
# num_pitches=num_pitches[voice_index],
# num_features_meta=left_metas.shape[-1],
# num_dense=num_dense, num_units_lstm=num_units_lstm,
# timesteps = timesteps)
model = deepBach_chord(num_features_lr=left_features.shape[-1],
num_features_c=central_features.shape[-1],
num_pitches=num_pitches[voice_index],
num_features_meta=left_metas.shape[-1],
num_dense=num_dense,
num_units_lstm=num_units_lstm,
timesteps=timesteps,
num_localseqs_lr=left_local_seqs.shape[-1])
elif 'skip' in model_name:
model = deepbach_skip_connections(
num_features_lr=left_features.shape[-1],
num_features_c=central_features.shape[-1],
num_features_meta=left_metas.shape[-1],
num_pitches=num_pitches[voice_index],
num_dense=num_dense,
num_units_lstm=num_units_lstm,
timesteps=timesteps)
else:
raise ValueError
model_path_name = 'models/' + model_name + '_' + str(voice_index)
if not os.path.exists(model_path_name + '.json') or create_new:
save_model(model, model_name=model_path_name, overwrite=create_new)