def main(args):
torch.cuda.manual_seed(13524532)
print("... Load trained models ...\n")
print(" Loding checkpoint of document-level TTS model: {}".format(
tts_ckpt))
print(" Loding checkpoint of MelGAN TTS model: {}".format(
args.mel_ckpt))
start = time.time()
mel_ckpt = torch.load(args.mel_ckpt)
if args.config is not None:
hp = HParam(args.config)
else:
hp = load_hparam_str(mel_ckpt['hp_str'])
model = Generator(hp.audio.n_mel_channels).cuda()
model.load_state_dict(mel_ckpt['model_g'])
model.eval(inference=False)
mel_time = time.time() - start
print('\n... Generate waveform ...\n')
with torch.no_grad():
num_of_iter = args.iteration
texts = []
with open(args.script_path, "r") as f:
for line in f:
line = line.strip()
if len(line):
texts.append(line)
print(" * input text\n {} \n".format(texts[0]))
for i in range(num_of_iter):
start = time.time()
mel, length, alignments = infer(args.tts_ckpt, texts[0])
if len(mel.shape) == 2:
mel = mel.unsqueeze(0)
mel = mel.cuda()
audio = model.inference(mel)
audio = audio.cpu().detach().numpy()
save_path = os.path.join(args.out_dir, str(i) + '_audio.wav')
write(save_path, hp.audio.sampling_rate, audio)
audio_length = len(audio) / hp.audio.sampling_rate
print(" {}. ".format(i + 1))
print(" - Path of generated audio file: {}".format(save_path))
print(" - Length of generated audio file: {}s".format(
audio_length))
print(
" - Time taken from text loading to generate spectrogram: : {}s"
.format(time.time() - start))
print(" - Time taken to generate waveform: : {}s\n".format(
time.time() - start + mel_time))
print("finished generation")
def main(args):
torch.manual_seed(1234)
torch.cuda.manual_seed(1234)
text_path = "/media/sh/Workspace/긴문장합성/kor_Document-level_Neural_TTS_length/test/1.txt"
doc_ckpt_kor = '/media/sh/Workspace/긴문장합성/kor_Document-level_Neural_TTS_length/outdir/checkpoint_29000'
save_folder = '/media/sh/Workspace/긴문장합성/samples'
today = datetime.datetime.today()
time = str(today.month) + str(today.day) + str(today.hour) + str(
today.minute) + str(today.second)
# time = str(time)
save_name = 'kor_audio_length_regul_' + doc_ckpt_kor.split(
'_')[-1] + '_' + time + '.wav'
save_path = os.path.join(save_folder, save_name)
checkpoint = torch.load(args.checkpoint_path)
if args.config is not None:
hp = HParam(args.config)
else:
hp = load_hparam_str(checkpoint['hp_str'])
model = Generator(hp.audio.n_mel_channels).cuda()
model.load_state_dict(checkpoint['model_g'])
model.eval(inference=False)
with torch.no_grad():
# for melpath in tqdm.tqdm(glob.glob(os.path.join(args.input_folder, '*.mel'))):
# mel = torch.load(melpath)
texts = []
with open(text_path, "r") as f:
for line in f:
line = line.strip()
if len(line):
texts.append(line)
mel, length, alignments = infer(doc_ckpt_kor, texts[0])
if len(mel.shape) == 2:
mel = mel.unsqueeze(0)
mel = mel.cuda()
audio = model.inference(mel)
audio = audio.cpu().detach().numpy()
# out_path = melpath.replace('.mel', '_reconstructed_epoch%04d.wav' % checkpoint['epoch'])
write(save_path, hp.audio.sampling_rate, audio)
print('합성 끝')
def get_melgan(full_path=None):
if not full_path:
melgan = torch.hub.load('seungwonpark/melgan', 'melgan')
melgan.eval()
melgan.to(device)
return melgan
# make sure to clone seungwonpark/melgan
print("use local vocoder")
from melgan.utils.hparams import load_hparam_str
from melgan.model.generator import Generator
cp = torch.load(full_path, map_location=device)
hp = load_hparam_str(cp["hp_str"])
model = Generator(hp.audio.n_mel_channels)
if torch.cuda.is_available():
model = model.cuda()
model.load_state_dict(cp["model_g"])
model.eval(inference=False)
model.to(device)
return model
def main(args):
checkpoint = torch.load(args.checkpoint_path)
if args.config is not None:
hp = HParam(args.config)
else:
hp = load_hparam_str(checkpoint['hp_str'])
model = Generator(hp.audio.n_mel_channels).cuda()
model.load_state_dict(checkpoint['model_g'])
model.eval(inference=False)
with torch.no_grad():
# for melpath in tqdm.tqdm(glob.glob(os.path.join(args.input_folder, '*.mel'))):
# mel = torch.load(melpath)
texts = []
with open("/media/qw/data/Experiment/Encoder_selfAtt/test/1.txt",
"r") as f:
for line in f:
line = line.strip()
if len(line):
texts.append(line)
for i in range(10):
mel, length, alignments = infer(
'/media/qw/data/Experiment/Encoder_selfAtt/tacotron2_statedict.pt',
texts[0])
# mel, length, alignments = infer('/media/qw/data/Experiment/Encoder_selfAtt/result/3sentence', texts[0])
# print('/'*i, '.'*(50-i))
# plt.figure()
# plt.imshow(alignments[0].T.cpu())
# plt.savefig('./align/alignment{}.png'.format(i), dpi=300)
# mel, length, alignments = infer('/media/qw/data/Experiment/Encoder_selfAtt/tacotron2_statedict.pt', 'Emil Sinclair is the protagonist of the novel. hello my name is sung woong hwang.')
if len(mel.shape) == 2:
mel = mel.unsqueeze(0)
mel = mel.cuda()
audio = model.inference(mel)
audio = audio.cpu().detach().numpy()
# out_path = melpath.replace('.mel', '_reconstructed_epoch%04d.wav' % checkpoint['epoch'])
write('/media/qw/data/Experiment/Encoder_selfAtt/audio.wav',
hp.audio.sampling_rate, audio)