def phase_3_train_encoder(params):
from io_modules.dataset import Dataset
from models.encoder import Encoder
from trainers.encoder import Trainer
trainset = Dataset("data/processed/train")
devset = Dataset("data/processed/dev")
sys.stdout.write('Found ' + str(len(trainset.files)) +
' training files and ' + str(len(devset.files)) +
' development files\n')
character2int = {}
for train_file in trainset.files:
from io_modules.dataset import DatasetIO
dio = DatasetIO()
lab_list = dio.read_lab(train_file + ".txt")
for entry in lab_list:
if entry.phoneme not in character2int:
character2int[entry.phoneme] = len(character2int)
sys.stdout.write('Found ' + str(len(character2int)) +
' unique phonemes\n')
f = open('data/models/encoder.chars', 'w')
for char in character2int:
f.write(
char.encode('utf-8') + '\t' + str(character2int[char]) + '\n')
f.close()
encoder = Encoder(params, len(character2int), character2int)
if params.resume:
sys.stdout.write('Resuming from previous checkpoint\n')
encoder.load('data/models/rnn_encoder')
trainer = Trainer(encoder, trainset, devset)
trainer.start_training(10, 1000)
def phase_4_train_pvocoder(params):
from io_modules.dataset import Dataset
from models.vocoder import Vocoder
from models.vocoder import ParallelVocoder
from trainers.vocoder import Trainer
vocoder_wavenet = Vocoder(params)
sys.stdout.write('Loading wavenet vocoder\n')
vocoder_wavenet.load('data/models/nn_vocoder')
vocoder = ParallelVocoder(params, vocoder_wavenet)
if params.resume:
sys.stdout.write('Resuming from previous checkpoint\n')
vocoder.load('data/models/pnn_vocoder')
trainset = Dataset("data/processed/train")
devset = Dataset("data/processed/dev")
sys.stdout.write('Found ' + str(len(trainset.files)) +
' training files and ' + str(len(devset.files)) +
' development files\n')
trainer = Trainer(vocoder,
trainset,
devset,
target_output_path='data/models/pnn_vocoder')
trainer.start_training(20,
params.batch_size,
params.target_sample_rate,
params=params)
def phase_3_train_encoder(params):
from io_modules.dataset import Dataset
from io_modules.dataset import Encodings
from models.encoder import Encoder
from trainers.encoder import Trainer
trainset = Dataset("data/processed/train")
devset = Dataset("data/processed/dev")
sys.stdout.write('Found ' + str(len(trainset.files)) +
' training files and ' + str(len(devset.files)) +
' development files\n')
encodings = Encodings()
count = 0
if not params.resume:
for train_file in trainset.files:
count += 1
if count % 100 == 0:
sys.stdout.write('\r' + str(count) + '/' +
str(len(trainset.files)) +
' processed files')
sys.stdout.flush()
from io_modules.dataset import DatasetIO
dio = DatasetIO()
lab_list = dio.read_lab(train_file + ".lab")
for entry in lab_list:
encodings.update(entry)
sys.stdout.write('\r' + str(count) + '/' +
str(len(trainset.files)) + ' processed files\n')
sys.stdout.write('Found ' + str(len(encodings.char2int)) +
' unique symbols, ' +
str(len(encodings.context2int)) +
' unique features and ' +
str(len(encodings.speaker2int)) +
' unique speakers\n')
encodings.store('data/models/encoder.encodings')
else:
encodings.load('data/models/encoder.encodings')
if params.resume:
runtime = True # avoid ortonormal initialization
else:
runtime = False
encoder = Encoder(params, encodings, runtime=runtime)
if params.resume:
sys.stdout.write('Resuming from previous checkpoint\n')
encoder.load('data/models/rnn_encoder')
if params.no_guided_attention:
sys.stdout.write('Disabling guided attention\n')
if params.no_bounds:
sys.stdout.write(
'Using internal stopping condition for synthesis\n')
trainer = Trainer(encoder, trainset, devset)
trainer.start_training(10, 1000, params)
def phase_5_test_vocoder(params):
from io_modules.dataset import Dataset
from models.vocoder import Vocoder
from trainers.vocoder import Trainer
vocoder = Vocoder(params, runtime=True)
vocoder.load('data/models/rnn')
trainset = Dataset("data/processed/train")
devset = Dataset("data/processed/dev")
sys.stdout.write('Found ' + str(len(trainset.files)) + ' training files and ' + str(
len(devset.files)) + ' development files\n')
trainer = Trainer(vocoder, trainset, devset)
trainer.synth_devset(params.batch_size, target_sample_rate=params.target_sample_rate, sample=True,
temperature=0.8)
def phase_2_train_vocoder(params):
from io_modules.dataset import Dataset
from models.vocoder import Vocoder
from trainers.vocoder import Trainer
vocoder = Vocoder(params)
if params.resume:
sys.stdout.write('Resuming from previous checkpoint\n')
vocoder.load('data/models/rnn')
trainset = Dataset("data/processed/train")
devset = Dataset("data/processed/dev")
sys.stdout.write('Found ' + str(len(trainset.files)) + ' training files and ' + str(
len(devset.files)) + ' development files\n')
trainer = Trainer(vocoder, trainset, devset)
trainer.start_training(20, params.batch_size, params.target_sample_rate)
def phase_7_train_sparse(params):
sys.stdout.write("Starting sparsification for VanillaLSTM\n")
from io_modules.dataset import Dataset
from models.vocoder import Vocoder
from trainers.vocoder import Trainer
vocoder = Vocoder(params, use_sparse_lstm=True)
sys.stdout.write('Resuming from previous checkpoint\n')
vocoder.load('data/models/rnn_vocoder_sparse')
sys.stdout.write("Reading datasets\n")
sys.stdout.flush()
trainset = Dataset("data/processed/train")
devset = Dataset("data/processed/dev")
sys.stdout.write('Found ' + str(len(trainset.files)) +
' training files and ' + str(len(devset.files)) +
' development files\n')
sys.stdout.flush()
trainer = Trainer(vocoder, trainset, devset)
trainer.start_training(20,
params.batch_size,
params.target_sample_rate,
params=params)