description=__doc__,
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument("--dataset", required=True, help='dataset name')
parser.add_argument("--warmstart",
help='Warmstart (transfer learn) from a pre-trained model')
args = parser.parse_args()
use_gpu = torch.cuda.is_available()
print('use_gpu', use_gpu)
if use_gpu:
torch.backends.cudnn.benchmark = True
if args.dataset not in ['ljspeech', 'mbspeech']:
from datasets.generic import vocab, Generic as SpeechDataset
train_data_loader = Text2MelDataLoader(text2mel_dataset=SpeechDataset(
['texts', 'mels', 'mel_gates'], args.dataset),
batch_size=64,
mode='train')
valid_data_loader = Text2MelDataLoader(text2mel_dataset=SpeechDataset(
['texts', 'mels', 'mel_gates'], args.dataset),
batch_size=64,
mode='valid')
else:
if args.dataset == 'ljspeech':
from datasets.lj_speech import vocab, LJSpeech as SpeechDataset
elif args.dataset == 'mbspeech':
from datasets.mb_speech import vocab, MBSpeech as SpeechDataset
train_data_loader = Text2MelDataLoader(text2mel_dataset=SpeechDataset(
['texts', 'mels', 'mel_gates']),
batch_size=64,
mode='train')
valid_data_loader = Text2MelDataLoader(text2mel_dataset=SpeechDataset(
def valid_dataloader(self, dataset, batch_size=32, \
num_workers=0 if sys.platform.startswith('win') else 8):
return Text2MelDataLoader(dataset,
batch_size=batch_size,
mode='valid',
num_workers=num_workers)
from datasets.swara import vocab, SWARA as SpeechDataset
elif args.dataset == 'swara_test':
from datasets.swara_test import vocab, SWARA as SpeechDataset
else:
print('No such dataset')
sys.exit(1)
# os.environ["CUDA_VISIBLE_DEVICES"]="3"
use_gpu = torch.cuda.is_available()
print('use_gpu', use_gpu)
if use_gpu:
torch.backends.cudnn.benchmark = True
train_data_loader = Text2MelDataLoader(text2mel_dataset=SpeechDataset(
['texts', 'mels', 'mel_gates', 'speakers', 'filenames']),
batch_size=16,
mode='train')
valid_data_loader = Text2MelDataLoader(text2mel_dataset=SpeechDataset(
['texts', 'mels', 'mel_gates', 'speakers', 'filenames']),
batch_size=16,
mode='valid')
text2mel = Text2Mel(vocab).cuda()
optimizer = torch.optim.Adam(text2mel.parameters(), lr=hp.text2mel_lr)
start_timestamp = int(time.time() * 1000)
start_epoch = 0
global_step = 0
logger = Logger(args.dataset, 'text2mel')
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument("--voice", default='Keira', help='voice name')
parser.add_argument("--script",
default='Keira_all.csv',
help='script filename')
args = parser.parse_args()
use_gpu = torch.cuda.is_available()
print('use_gpu', use_gpu)
if use_gpu:
torch.backends.cudnn.benchmark = True
Speech.load(['texts', 'mels', 'mel_gates'], args.voice, args.script)
train_data_loader = Text2MelDataLoader(Speech,
batch_size=hp.text2mel_batch_size,
mode='train')
valid_data_loader = Text2MelDataLoader(Speech,
batch_size=hp.text2mel_batch_size,
mode='valid')
text2mel = Text2Mel(vocab).cuda()
optimizer = torch.optim.Adam(text2mel.parameters(), lr=hp.text2mel_lr)
start_timestamp = int(time.time() * 1000)
start_epoch = 0
global_step = 0
logger = Logger(
f'{args.voice}-{args.script}-{hp.d}-{hp.text2mel_lr}-{hp.text2mel_batch_size}',
else:
index = 'cpu'
device = select_device(index)
hparams = HParam(args.config) \
if args.config else HParam(osp.join(osp.abspath(os.getcwd()), 'config', 'default.yaml'))
checkpoint = args.checkpoint or get_last_chkpt_path(
osp.join(hparams.trainer.logdir, f"{hparams.data.dataset}-{args.name}"))
extractor = DurationTrainer(hparams, device=device).load_checkpoint(checkpoint).model
extractor.train(False)
dataset_root = osp.join(hparams.data.datasets_path, hparams.data.dataset_dir)
dataset = SpeechDataset(['mels', 'mlens', 'texts', 'tlens', 'files'], dataset_root, hparams.text)
dataloader = Text2MelDataLoader(dataset, args.batch_size, mode='whole')
normalizer = MinMaxNorm(hparams.audio.spec_min, hparams.audio.spec_max)
pbar = tqdm(dataloader, unit="audios", unit_scale=dataloader.batch_size, \
disable=hparams.trainer.disable_progress_bar)
with open(osp.join(dataset.path, 'duration.txt'), 'w', encoding='utf-8') as fw:
for it, batch in enumerate(pbar, start=1):
mels, mlens, texts, tlens = \
batch['mels'], batch['mlens'].squeeze(1), batch['texts'].long(), batch['tlens'].squeeze(1)
mels, mlens, texts, tlens = \
mels.to(device), mlens.to(device), texts.to(device), tlens.to(device)
mels = normalizer(mels)
with torch.no_grad():
melspecs, attns = extractor((texts, tlens, mels, True))