def start_training(self,
itt_no_improve,
batch_size,
target_sample_rate,
params=None):
epoch = 1
left_itt = itt_no_improve
dio = DatasetIO()
self._render_devset()
sys.stdout.write("\n")
# self.synth_devset(batch_size, target_sample_rate)
self.vocoder.store(self.target_output_path)
num_files = 0
while left_itt > 0:
sys.stdout.write("Starting epoch " + str(epoch) + "\n")
sys.stdout.write("Shuffling training data\n")
from random import shuffle
shuffle(self.trainset.files)
file_index = 1
total_loss = 0
for file in self.trainset.files:
num_files += 1
sys.stdout.write("\t" + str(file_index) + "/" +
str(len(self.trainset.files)) +
" processing file " + file + '\n')
sys.stdout.flush()
wav_file = file + ".orig.wav"
mgc_file = file + ".mgc.npy"
mgc = np.load(mgc_file)
file_index += 1
data, sample_rate = dio.read_wave(wav_file)
# wave_disc = data * 32768
wave_disc = np.array(data, dtype=np.float32)
import time
start = time.time()
loss = self.vocoder.learn(wave_disc, mgc, batch_size)
total_loss += loss
stop = time.time()
sys.stdout.write(' avg loss=' + str(loss) +
" execution time=" + str(stop - start))
sys.stdout.write('\n')
sys.stdout.flush()
if file_index % 5000 == 0:
self.vocoder.store(self.target_output_path)
self.synth_devset(batch_size, target_sample_rate)
self.vocoder.store(self.target_output_path)
self.synth_devset(batch_size, target_sample_rate)
epoch += 1
def start_training(self, itt_no_improve, batch_size, target_sample_rate):
epoch = 1
left_itt = itt_no_improve
dio = DatasetIO()
self._render_devset()
sys.stdout.write("\n")
self.vocoder.store('data/models/rnn_vocoder')
while left_itt > 0:
sys.stdout.write("Starting epoch " + str(epoch) + "\n")
sys.stdout.write("Shuffling training data\n")
from random import shuffle
shuffle(self.trainset.files)
file_index = 1
total_loss = 0
for file in self.trainset.files:
sys.stdout.write("\t" + str(file_index) + "/" +
str(len(self.trainset.files)) +
" processing file " + file)
sys.stdout.flush()
wav_file = file + ".orig.wav"
mgc_file = file + ".mgc.npy"
mgc = np.load(mgc_file)
file_index += 1
data, sample_rate = dio.read_wave(wav_file)
if self.use_ulaw:
[wave_disc, ulaw_cont] = dio.ulaw_encode(data)
else:
wave_disc = dio.b16_enc(data)
import time
start = time.time()
loss = self.vocoder.learn(wave_disc, mgc, batch_size)
total_loss += loss
stop = time.time()
sys.stdout.write(' avg loss=' + str(loss) +
" execution time=" + str(stop - start))
sys.stdout.write('\n')
sys.stdout.flush()
if file_index % 50 == 0:
self.synth_devset(batch_size, target_sample_rate)
self.vocoder.store('data/models/rnn_vocoder')
self.synth_devset(batch_size, target_sample_rate)
self.vocoder.store('data/models/rnn_vocoder')
epoch += 1
def phase_1_prepare_corpus(params):
from os import listdir
from os.path import isfile, join
from os.path import exists
train_files_tmp = [
f for f in listdir(params.train_folder)
if isfile(join(params.train_folder, f))
]
dev_files_tmp = [
f for f in listdir(params.dev_folder)
if isfile(join(params.dev_folder, f))
]
sys.stdout.write("Scanning training files...")
sys.stdout.flush()
final_list = []
for file in train_files_tmp:
base_name = file[:-4]
lab_name = base_name + '.txt'
wav_name = base_name + '.wav'
if exists(join(params.train_folder, lab_name)) and exists(
join(params.train_folder, wav_name)):
if base_name not in final_list:
final_list.append(base_name)
train_files = final_list
sys.stdout.write(" found " + str(len(train_files)) +
" valid training files\n")
sys.stdout.write("Scanning development files...")
sys.stdout.flush()
final_list = []
for file in dev_files_tmp:
base_name = file[:-4]
lab_name = base_name + '.txt'
wav_name = base_name + '.wav'
if exists(join(params.dev_folder, lab_name)) and exists(
join(params.dev_folder, wav_name)):
if base_name not in final_list:
final_list.append(base_name)
dev_files = final_list
sys.stdout.write(" found " + str(len(dev_files)) +
" valid development files\n")
from io_modules.dataset import DatasetIO
from io_modules.vocoder import MelVocoder
from shutil import copyfile
import pysptk
dio = DatasetIO()
vocoder = MelVocoder()
base_folder = params.train_folder
for index in range(len(train_files)):
sys.stdout.write("\r\tprocessing file " + str(index + 1) + "/" +
str(len(train_files)))
sys.stdout.flush()
base_name = train_files[index]
txt_name = base_name + '.txt'
wav_name = base_name + '.wav'
spc_name = base_name + '.png'
lab_name = base_name + '.lab'
# LAB - copy or create
if exists(join(base_folder, lab_name)):
copyfile(join(base_folder, lab_name),
join('data/processed/train', lab_name))
else:
create_lab_file(join(base_folder, txt_name),
join('data/processed/train', lab_name))
# TXT
copyfile(join(base_folder, txt_name),
join('data/processed/train', txt_name))
# WAVE
data, sample_rate = dio.read_wave(
join(base_folder, wav_name),
sample_rate=params.target_sample_rate)
mgc = vocoder.melspectrogram(data,
sample_rate=params.target_sample_rate,
num_mels=params.mgc_order)
# SPECT
render_spectrogram(mgc, join('data/processed/train', spc_name))
dio.write_wave(
join('data/processed/train', base_name + '.orig.wav'), data,
sample_rate)
array2file(mgc, join('data/processed/train', base_name + '.mgc'))
sys.stdout.write('\n')
base_folder = params.dev_folder
for index in range(len(dev_files)):
sys.stdout.write("\r\tprocessing file " + str(index + 1) + "/" +
str(len(dev_files)))
sys.stdout.flush()
base_name = dev_files[index]
txt_name = base_name + '.txt'
wav_name = base_name + '.wav'
spc_name = base_name + '.png'
lab_name = base_name + '.lab'
# LAB - copy or create
if exists(join(base_folder, lab_name)):
copyfile(join(base_folder, lab_name),
join('data/processed/dev', lab_name))
else:
create_lab_file(join(base_folder, txt_name),
join('data/processed/dev', lab_name))
# TXT
copyfile(join(base_folder, txt_name),
join('data/processed/dev/', txt_name))
# WAVE
data, sample_rate = dio.read_wave(
join(base_folder, wav_name),
sample_rate=params.target_sample_rate)
mgc = vocoder.melspectrogram(data,
sample_rate=params.target_sample_rate,
num_mels=params.mgc_order)
# SPECT
render_spectrogram(mgc, join('data/processed/dev', spc_name))
dio.write_wave(join('data/processed/dev', base_name + '.orig.wav'),
data, sample_rate)
array2file(mgc, join('data/processed/dev', base_name + '.mgc'))
sys.stdout.write('\n')
def start_training(self,
itt_no_improve,
batch_size,
target_sample_rate,
params=None):
epoch = 1
left_itt = itt_no_improve
dio = DatasetIO()
self._render_devset()
sys.stdout.write("\n")
if self.vocoder.sparse:
print("Setting sparsity at: " + str(params.sparsity_step) + "%")
sparsity = params.sparsity_step
self.vocoder.rnnFine.set_sparsity(float(sparsity) / 100)
self.vocoder.rnnCoarse.set_sparsity(float(sparsity) / 100)
if self.vocoder.sparse:
self.vocoder.store('data/models/rnn_vocoder_sparse')
else:
self.vocoder.store('data/models/rnn_vocoder')
num_files = 0
while left_itt > 0:
sys.stdout.write("Starting epoch " + str(epoch) + "\n")
sys.stdout.write("Shuffling training data\n")
from random import shuffle
shuffle(self.trainset.files)
file_index = 1
total_loss = 0
for file in self.trainset.files:
num_files += 1
if num_files == params.sparsity_increase:
sparsity += params.sparsity_step
num_files = 0
if sparsity <= params.sparsity_target:
print("Setting sparsity at " + str(sparsity) + "%")
self.vocoder.rnnFine.set_sparsity(
float(sparsity) / 100)
self.vocoder.rnnCoarse.set_sparsity(
float(sparsity) / 100)
else:
sparsity = params.sparsity_target
sys.stdout.write("\t" + str(file_index) + "/" +
str(len(self.trainset.files)) +
" processing file " + file)
sys.stdout.flush()
wav_file = file + ".orig.wav"
mgc_file = file + ".mgc.npy"
mgc = np.load(mgc_file)
file_index += 1
data, sample_rate = dio.read_wave(wav_file)
if self.use_ulaw:
[wave_disc, ulaw_cont] = dio.ulaw_encode(data)
else:
wave_disc = dio.b16_enc(data)
import time
start = time.time()
loss = self.vocoder.learn(wave_disc, mgc, batch_size)
total_loss += loss
stop = time.time()
sys.stdout.write(' avg loss=' + str(loss) +
" execution time=" + str(stop - start))
sys.stdout.write('\n')
sys.stdout.flush()
if file_index % 50 == 0:
self.synth_devset(batch_size, target_sample_rate)
if self.vocoder.sparse:
self.vocoder.store('data/models/rnn_vocoder_sparse')
else:
self.vocoder.store('data/models/rnn_vocoder')
self.synth_devset(batch_size, target_sample_rate)
if self.vocoder.sparse:
self.vocoder.store('data/models/rnn_vocoder_sparse')
else:
self.vocoder.store('data/models/rnn_vocoder')
epoch += 1
def phase_1_prepare_corpus(params):
from os import listdir
from os.path import isfile, join
from os.path import exists
train_files_tmp = [
f for f in listdir(params.train_folder)
if isfile(join(params.train_folder, f))
]
if params.dev_folder is not None:
dev_files_tmp = [
f for f in listdir(params.dev_folder)
if isfile(join(params.dev_folder, f))
]
else:
dev_files_tmp = []
if params.g2p is not None:
from models.g2p import G2P
from io_modules.encodings import Encodings
g2p_encodings = Encodings()
g2p_encodings.load(params.g2p + '.encodings')
g2p = G2P(g2p_encodings)
g2p.load(params.g2p + '-bestAcc.network')
if exists(params.g2p + '.lexicon'):
g2p.load_lexicon(params.g2p + '.lexicon')
else:
g2p = None
sys.stdout.write("Scanning training files...")
sys.stdout.flush()
final_list = []
for file in train_files_tmp:
base_name = file[:-4]
lab_name = base_name + '.txt'
wav_name = base_name + '.wav'
if exists(join(params.train_folder, lab_name)) and exists(
join(params.train_folder, wav_name)):
if base_name not in final_list:
final_list.append(base_name)
train_files = final_list
sys.stdout.write(" found " + str(len(train_files)) +
" valid training files\n")
sys.stdout.write("Scanning development files...")
sys.stdout.flush()
final_list = []
for file in dev_files_tmp:
base_name = file[:-4]
lab_name = base_name + '.txt'
wav_name = base_name + '.wav'
if exists(join(params.dev_folder, lab_name)) and exists(
join(params.dev_folder, wav_name)):
if base_name not in final_list:
final_list.append(base_name)
dev_files = final_list
sys.stdout.write(" found " + str(len(dev_files)) +
" valid development files\n")
from io_modules.dataset import DatasetIO
from io_modules.vocoder import MelVocoder
from shutil import copyfile
dio = DatasetIO()
vocoder = MelVocoder()
base_folder = params.train_folder
total_files = 0
for index in range(len(train_files)):
total_files += 1
sys.stdout.write("\r\tprocessing file " + str(index + 1) + "/" +
str(len(train_files)))
sys.stdout.flush()
base_name = train_files[index]
txt_name = base_name + '.txt'
wav_name = base_name + '.wav'
spc_name = base_name + '.png'
lab_name = base_name + '.lab'
tgt_txt_name = txt_name
tgt_spc_name = spc_name
tgt_lab_name = lab_name
if params.prefix is not None:
tgt_txt_name = params.prefix + "_{:05d}".format(
total_files) + '.txt'
tgt_spc_name = params.prefix + "_{:05d}".format(
total_files) + '.png'
tgt_lab_name = params.prefix + "_{:05d}".format(
total_files) + '.lab'
# LAB - copy or create
if exists(join(base_folder, lab_name)):
copyfile(join(base_folder, lab_name),
join('data/processed/train', tgt_lab_name))
else:
create_lab_file(join(base_folder, txt_name),
join('data/processed/train', tgt_lab_name),
speaker_name=params.speaker,
g2p=g2p)
# TXT
copyfile(join(base_folder, txt_name),
join('data/processed/train', tgt_txt_name))
# WAVE
data, sample_rate = dio.read_wave(
join(base_folder, wav_name),
sample_rate=params.target_sample_rate)
mgc = vocoder.melspectrogram(data,
sample_rate=params.target_sample_rate,
num_mels=params.mgc_order)
# SPECT
render_spectrogram(mgc, join('data/processed/train', tgt_spc_name))
if params.prefix is None:
dio.write_wave(
join('data/processed/train', base_name + '.orig.wav'),
data, sample_rate)
array2file(mgc, join('data/processed/train',
base_name + '.mgc'))
else:
tgt_wav_name = params.prefix + "_{:05d}".format(
total_files) + '.orig.wav'
tgt_mgc_name = params.prefix + "_{:05d}".format(
total_files) + '.mgc'
dio.write_wave(join('data/processed/train', tgt_wav_name),
data, sample_rate)
array2file(mgc, join('data/processed/train', tgt_mgc_name))
sys.stdout.write('\n')
base_folder = params.dev_folder
for index in range(len(dev_files)):
total_files += 1
sys.stdout.write("\r\tprocessing file " + str(index + 1) + "/" +
str(len(dev_files)))
sys.stdout.flush()
base_name = dev_files[index]
txt_name = base_name + '.txt'
wav_name = base_name + '.wav'
spc_name = base_name + '.png'
lab_name = base_name + '.lab'
tgt_txt_name = txt_name
tgt_spc_name = spc_name
tgt_lab_name = lab_name
if params.prefix is not None:
tgt_txt_name = params.prefix + "_{:05d}".format(
total_files) + '.txt'
tgt_spc_name = params.prefix + "_{:05d}".format(
total_files) + '.png'
tgt_lab_name = params.prefix + "_{:05d}".format(
total_files) + '.lab'
# LAB - copy or create
if exists(join(base_folder, lab_name)):
copyfile(join(base_folder, lab_name),
join('data/processed/dev', tgt_lab_name))
else:
create_lab_file(join(base_folder, txt_name),
join('data/processed/dev', tgt_lab_name),
speaker_name=params.speaker,
g2p=g2p)
# TXT
copyfile(join(base_folder, txt_name),
join('data/processed/dev', tgt_txt_name))
# WAVE
data, sample_rate = dio.read_wave(
join(base_folder, wav_name),
sample_rate=params.target_sample_rate)
mgc = vocoder.melspectrogram(data,
sample_rate=params.target_sample_rate,
num_mels=params.mgc_order)
# SPECT
render_spectrogram(mgc, join('data/processed/dev', tgt_spc_name))
if params.prefix is None:
dio.write_wave(
join('data/processed/dev', base_name + '.orig.wav'), data,
sample_rate)
array2file(mgc, join('data/processed/dev', base_name + '.mgc'))
else:
tgt_wav_name = params.prefix + "_{:05d}".format(
total_files) + '.orig.wav'
tgt_mgc_name = params.prefix + "_{:05d}".format(
total_files) + '.mgc'
dio.write_wave(join('data/processed/dev', tgt_wav_name), data,
sample_rate)
array2file(mgc, join('data/processed/dev', tgt_mgc_name))
sys.stdout.write('\n')