def run():
parser = argparse.ArgumentParser()
parser.add_argument('--caching_dir',
default='/datasets/models/tacotron/cache')
parser.add_argument(
'--hparams',
default='',
help=
'Hyperparameter overrides as a comma-separated list of name=value pairs'
)
accepted_modes = ['eval', 'synthesis', 'live']
parser.add_argument(
'--mode',
default='eval',
help='mode of run: can be one of {}'.format(accepted_modes))
parser.add_argument(
'--GTA',
default='True',
help=
'Ground truth aligned synthesis, defaults to True, only considered in synthesis mode'
)
parser.add_argument(
'--text_list',
default='',
help=
'Text file contains list of texts to be synthesized. Valid if mode=eval'
)
parser.add_argument(
'--speaker_id',
default=None,
help=
'Defines the speakers ids to use when running standalone Wavenet on a folder of mels. this variable must be a comma-separated list of ids'
)
args = parser.parse_args()
if args.mode not in accepted_modes:
raise ValueError('accepted modes are: {}, found {}'.format(
accepted_modes, args.mode))
if args.GTA not in ('True', 'False'):
raise ValueError('GTA option must be either True or False')
modified_hp = hparams.parse(args.hparams)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
from src.tac.training.wav_training import get_log_dir
from src.tac.training.wav_training import get_save_dir
wavenet_log_dir = get_log_dir(args.caching_dir)
wavenet_pretrained = get_save_dir(wavenet_log_dir)
log('Synthesizing audio from mel-spectrograms.. (This may take a while)')
run_synthesis(args, wavenet_pretrained, args.caching_dir, hparams)
log('Tacotron-2 TTS synthesis complete!')
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--base_dir', default='')
parser.add_argument(
'--hparams',
default='',
help=
'Hyperparameter overrides as a comma-separated list of name=value pairs'
)
parser.add_argument('--metadata', default='tacotron_output/gta/map.txt')
args = parser.parse_args()
modified_hparams = hparams.parse(args.hparams)
run(args, modified_hparams)
def run():
from multiprocessing import cpu_count
from src.tac.preprocessing.Dataset import LJSPEECH_TEST, LJSPEECH_LITE, LJSPEECH
from src.tac.hparams import hparams
print('initializing preprocessing..')
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', default=LJSPEECH)
parser.add_argument('--cache_path', default='/datasets/models/tacotron/cache')
parser.add_argument('--n_jobs', type=int, default=cpu_count())
parser.add_argument('--hparams', default='', help='Hyperparameter overrides as a comma-separated list of name=value pairs')
args = parser.parse_args()
modified_hp = hparams.parse(args.hparams)
processor = Preprocessor(args.n_jobs, args.cache_path, args.dataset, modified_hp)
processor.run()
def main():
print('initializing preprocessing..')
parser = argparse.ArgumentParser()
parser.add_argument('--base_dir', default='')
parser.add_argument(
'--hparams',
default='',
help=
'Hyperparameter overrides as a comma-separated list of name=value pairs'
)
parser.add_argument('--input_dir', default='LJSpeech-1.1/wavs')
parser.add_argument('--output', default='tacotron_output/gta/')
parser.add_argument('--n_jobs', type=int, default=cpu_count())
args = parser.parse_args()
modified_hp = hparams.parse(args.hparams)
run_preprocess(args, modified_hp)
def run():
parser = argparse.ArgumentParser()
parser.add_argument('--caching_dir', default='/datasets/models/tacotron/cache')
parser.add_argument('--mode', default='synthesis', help='mode for synthesis of tacotron after training')
parser.add_argument('--GTA', default='True', help='Ground truth aligned synthesis, defaults to True, only considered in Tacotron synthesis mode')
parser.add_argument('--tf_log_level', type=int, default=1, help='Tensorflow C++ log level.')
parser.add_argument('--hparams', default='', help='Hyperparameter overrides as a comma-separated list of name=value pairs')
args = parser.parse_args()
modified_hp = hparams.parse(args.hparams)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = str(args.tf_log_level)
log_dir = get_log_dir(args.caching_dir)
os.makedirs(log_dir, exist_ok=True)
infolog_path = get_infolog_path(log_dir)
infolog.init(infolog_path, 'tacotron')
log('\n##########################################################\n')
log('Tacotron GTA Synthesis\n')
log('###########################################################\n')
tacotron_pretrained = get_save_dir(log_dir)
run_synthesis(args, tacotron_pretrained, modified_hp)
def run():
parser = argparse.ArgumentParser()
parser.add_argument('--caching_dir',
default='/datasets/models/tacotron/cache')
parser.add_argument('--checkpoint',
default='pretrained/',
help='Path to model checkpoint')
parser.add_argument(
'--hparams',
default='',
help=
'Hyperparameter overrides as a comma-separated list of name=value pairs'
)
#parser.add_argument('--name', help='Name of logging directory if the two models were trained together.')
#parser.add_argument('--tacotron_name', help='Name of logging directory of Tacotron. If trained separately')
#parser.add_argument('--wavenet_name', help='Name of logging directory of WaveNet. If trained separately')
#parser.add_argument('--model', default='Tacotron-2')
#parser.add_argument('--input_dir', default='training_data/', help='folder to contain inputs sentences/targets')
#parser.add_argument('--mels_dir', default='tacotron_output/eval/', help='folder to contain mels to synthesize audio from using the Wavenet')
#parser.add_argument('--output_dir', default='output/', help='folder to contain synthesized mel spectrograms')
accepted_modes = ['eval', 'synthesis', 'live']
parser.add_argument(
'--mode',
default='eval',
help='mode of run: can be one of {}'.format(accepted_modes))
parser.add_argument(
'--GTA',
default='True',
help=
'Ground truth aligned synthesis, defaults to True, only considered in synthesis mode'
)
parser.add_argument(
'--text_list',
default='',
help=
'Text file contains list of texts to be synthesized. Valid if mode=eval'
)
parser.add_argument(
'--speaker_id',
default=None,
help=
'Defines the speakers ids to use when running standalone Wavenet on a folder of mels. this variable must be a comma-separated list of ids'
)
args = parser.parse_args()
if args.mode not in accepted_modes:
raise ValueError('accepted modes are: {}, found {}'.format(
accepted_modes, args.mode))
if args.GTA not in ('True', 'False'):
raise ValueError('GTA option must be either True or False')
modified_hp = hparams.parse(args.hparams)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
from src.tac.training.tacotron_training import get_log_dir
from src.tac.training.tacotron_training import get_save_dir
taco_log_dir = get_log_dir(args.caching_dir)
tacotron_pretrained = get_save_dir(taco_log_dir)
#run_name = args.name or args.tacotron_name or args.model
#taco_checkpoint = os.path.join('logs-' + run_name, 'taco_' + args.checkpoint)
sentences = get_sentences(args)
log('Synthesizing mel-spectrograms from text..')
run_eval(args, tacotron_pretrained, modified_hp, sentences)
self.processing_result = result
return result
def show_stats(self):
assert self.processing_result
textlenght_sum = sum([int(m[1]) for m in self.processing_result])
textlenght_max = max([int(m[1]) for m in self.processing_result])
print('Written {} utterances'.format(len(self.processing_result)))
print('Sum input length (text chars): {}'.format(textlenght_sum))
print('Max input length (text chars): {}'.format(textlenght_max))
if __name__ == "__main__":
from hparams import hparams
from multiprocessing import cpu_count
if __name__ == "__main__":
from src.preprocessing.parser.LJSpeechDatasetParser import LJSpeechDatasetParser
from src.preprocessing.parser.DummyIPADatasetParser import DummyIPADatasetParser
#parser = LJSpeechDatasetParser('/datasets/LJSpeech-1.1-test')
parser = DummyIPADatasetParser('/datasets/IPA-Dummy')
processor = TextProcessor(hparams.parse(''),
'/datasets/models/tacotron/cache')
processor.process(parser, cpu_count())
processor.show_stats()
def run(testrun: bool = False):
import argparse
import os
import tensorflow as tf
from src.tac import infolog
from src.tac.hparams import hparams
parser = argparse.ArgumentParser()
train_steps = 20000
checkpoint_intervall = 1000
if testrun:
train_steps = 10
checkpoint_intervall = 1
parser.add_argument('--caching_dir',
default='/datasets/models/tacotron/cache')
parser.add_argument('--tacotron_train_steps',
type=int,
default=train_steps,
help='total number of tacotron training steps')
parser.add_argument('--tf_log_level',
type=int,
default=1,
help='Tensorflow C++ log level.')
parser.add_argument('--checkpoint_interval',
type=int,
default=checkpoint_intervall,
help='Steps between writing checkpoints') # 2500
parser.add_argument(
'--hparams',
default='',
help=
'Hyperparameter overrides as a comma-separated list of name=value pairs'
)
parser.add_argument('--restore',
type=bool,
default=False,
help='Set this to False to do a fresh training')
parser.add_argument('--eval_interval',
type=int,
default=30000,
help='Steps between eval on test data')
parser.add_argument('--summary_interval',
type=int,
default=30000,
help='Steps between running summary ops')
parser.add_argument(
'--embedding_interval',
type=int,
default=30000,
help='Steps between updating embeddings projection visualization')
args = parser.parse_args()
modified_hp = hparams.parse(args.hparams)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = str(args.tf_log_level)
log_dir = get_log_dir(args.caching_dir)
os.makedirs(log_dir, exist_ok=True)
infolog_path = get_infolog_path(log_dir)
infolog.init(infolog_path, 'tacotron')
log('\n##########################################################\n')
log('Tacotron Train\n')
log('###########################################################\n')
train(log_dir, args, modified_hp)