def load(self, t_checkpoint_path, v_checkpoint_path,
t_config_path=None, v_config_path=None, model_name='tacotron'):
if t_checkpoint_path.endswith('.pt'):
self.model_name = 'nvidia'
print('Constructing model: %s' % self.model_name)
# set-up params
hparams = create_hparams()
# load model from checkpoint
self.model = Tacotron2(hparams)
self.model.load_state_dict(torch.load(t_checkpoint_path,
map_location='cpu')['state_dict'])
_ = self.model.eval()
else: # elif t_checkpoint_path.endswith('.pth.tar'):
self.model_name = 'coqui'
print('Constructing model: %s' % self.model_name)
# load tts config and audio processor
self.tts_config = load_config(t_config_path)
self.tts_model = setup_tts_model(config=self.tts_config)
self.tts_model.load_checkpoint(self.tts_config,
t_checkpoint_path, eval=True)
self.ap = AudioProcessor(verbose=False, **self.tts_config.audio)
# load vocoder config and audio processor
vocoder_config = load_config(v_config_path)
self.vocoder_ap = AudioProcessor(verbose=False, **vocoder_config.audio)
# Load neurips MelGAN for mel2audio synthesis
self.vocoder = torch.hub.load('descriptinc/melgan-neurips', 'load_melgan')
melgan_ckpt = torch.load(v_checkpoint_path, map_location='cpu')
self.vocoder.mel2wav.load_state_dict(melgan_ckpt)
def _load_tts(self, tts_checkpoint: str, tts_config_path: str, use_cuda: bool) -> None:
"""Load the TTS model.
1. Load the model config.
2. Init the model from the config.
3. Load the model weights.
4. Move the model to the GPU if CUDA is enabled.
5. Init the speaker manager in the model.
Args:
tts_checkpoint (str): path to the model checkpoint.
tts_config_path (str): path to the model config file.
use_cuda (bool): enable/disable CUDA use.
"""
# pylint: disable=global-statement
self.tts_config = load_config(tts_config_path)
if self.tts_config["use_phonemes"] and self.tts_config["phonemizer"] is None:
raise ValueError("Phonemizer is not defined in the TTS config.")
self.tts_model = setup_tts_model(config=self.tts_config)
if not self.encoder_checkpoint:
self._set_speaker_encoder_paths_from_tts_config()
self.tts_model.load_checkpoint(self.tts_config, tts_checkpoint, eval=True)
if use_cuda:
self.tts_model.cuda()
if self.encoder_checkpoint and hasattr(self.tts_model, "speaker_manager"):
self.tts_model.speaker_manager.init_encoder(self.encoder_checkpoint, self.encoder_config, use_cuda)
def _load_tts(self, tts_checkpoint: str, tts_config_path: str,
use_cuda: bool) -> None:
"""Load the TTS model.
1. Load the model config.
2. Init the AudioProcessor.
3. Init the model from the config.
4. Move the model to the GPU if CUDA is enabled.
5. Init the speaker manager for the model.
Args:
tts_checkpoint (str): path to the model checkpoint.
tts_config_path (str): path to the model config file.
use_cuda (bool): enable/disable CUDA use.
"""
# pylint: disable=global-statement
self.tts_config = load_config(tts_config_path)
self.use_phonemes = self.tts_config.use_phonemes
self.ap = AudioProcessor(verbose=False, **self.tts_config.audio)
speaker_manager = self._init_speaker_manager()
language_manager = self._init_language_manager()
if not self.encoder_checkpoint:
self._set_speaker_encoder_paths_from_tts_config()
speaker_manager = self._init_speaker_encoder(speaker_manager)
if language_manager is not None:
self.tts_model = setup_tts_model(
config=self.tts_config,
speaker_manager=speaker_manager,
language_manager=language_manager,
)
else:
self.tts_model = setup_tts_model(config=self.tts_config,
speaker_manager=speaker_manager)
self.tts_model.load_checkpoint(self.tts_config,
tts_checkpoint,
eval=True)
if use_cuda:
self.tts_model.cuda()
def _load_tts(self, tts_checkpoint: str, tts_config_path: str,
use_cuda: bool) -> None:
"""Load the TTS model.
Args:
tts_checkpoint (str): path to the model checkpoint.
tts_config_path (str): path to the model config file.
use_cuda (bool): enable/disable CUDA use.
"""
# pylint: disable=global-statement
self.tts_config = load_config(tts_config_path)
self.use_phonemes = self.tts_config.use_phonemes
self.ap = AudioProcessor(verbose=False, **self.tts_config.audio)
self.tts_model = setup_tts_model(config=self.tts_config)
self.tts_model.load_checkpoint(self.tts_config,
tts_checkpoint,
eval=True)
if use_cuda:
self.tts_model.cuda()
self._set_tts_speaker_file()
def _load_tts(self, tts_checkpoint: str, tts_config_path: str,
use_cuda: bool) -> None:
"""Load the TTS model.
Args:
tts_checkpoint (str): path to the model checkpoint.
tts_config_path (str): path to the model config file.
use_cuda (bool): enable/disable CUDA use.
"""
# pylint: disable=global-statement
self.tts_config = load_config(tts_config_path)
# Patch stats_path
stats_path = self.tts_config["audio"].get("stats_path", "")
if stats_path and (not os.path.isfile(stats_path)):
stats_path = os.path.join(os.path.dirname(tts_checkpoint),
os.path.split(stats_path)[1])
self.tts_config["audio"]["stats_path"] = stats_path
# Patch speakers file
speakers_file = self.tts_config.get("model_args",
{}).get("speakers_file", "")
if speakers_file and (not os.path.isfile(speakers_file)):
speakers_file = os.path.join(os.path.dirname(tts_checkpoint),
os.path.split(speakers_file)[1])
self.tts_config["model_args"]["speakers_file"] = speakers_file
self.use_phonemes = self.tts_config.use_phonemes
self.ap = AudioProcessor(verbose=False, **self.tts_config.audio)
self.tts_model = setup_tts_model(config=self.tts_config)
self.tts_model.load_checkpoint(self.tts_config,
tts_checkpoint,
eval=True)
if use_cuda:
self.tts_model.cuda()
self._set_tts_speaker_file()