def init_multispeaker(self, config: Coqpit, data: List = None):
"""Initialize multi-speaker modules of a model. A model can be trained either with a speaker embedding layer
or with external `d_vectors` computed from a speaker encoder model.
If you need a different behaviour, override this function for your model.
Args:
config (Coqpit): Model configuration.
data (List, optional): Dataset items to infer number of speakers. Defaults to None.
"""
if hasattr(config, "model_args"):
config = config.model_args
self.embedded_speaker_dim = 0
# init speaker manager
self.speaker_manager = get_speaker_manager(config, data=data)
if config.num_speakers > 0 and self.speaker_manager.num_speakers == 0:
self.speaker_manager.num_speakers = config.num_speakers
self.num_speakers = self.speaker_manager.num_speakers
# init speaker embedding layer
if config.use_speaker_embedding and not config.use_d_vector_file:
self.embedded_speaker_dim = config.speaker_embedding_channels
self.emb_g = nn.Embedding(config.num_speakers,
config.speaker_embedding_channels)
# init d-vector usage
if config.use_d_vector_file:
self.embedded_speaker_dim = config.d_vector_dim
def main(args): # pylint: disable=redefined-outer-name
# pylint: disable=global-variable-undefined
global meta_data, speaker_manager
# Audio processor
ap = AudioProcessor(**c.audio)
# load data instances
meta_data_train, meta_data_eval = load_meta_data(c.datasets,
eval_split=args.eval)
# use eval and training partitions
meta_data = meta_data_train + meta_data_eval
# parse speakers
speaker_manager = get_speaker_manager(c, args, meta_data_train)
# setup model
model = setup_model(c)
# restore model
model.load_checkpoint(c, args.checkpoint_path, eval=True)
if use_cuda:
model.cuda()
num_params = count_parameters(model)
print("\n > Model has {} parameters".format(num_params), flush=True)
# set r
r = 1 if c.model.lower() == "glow_tts" else model.decoder.r
own_loader = setup_loader(ap, r, verbose=True)
extract_spectrograms(
own_loader,
model,
ap,
args.output_path,
quantized_wav=args.quantized,
save_audio=args.save_audio,
debug=args.debug,
metada_name="metada.txt",
)
def init_multispeaker(self, config: Coqpit, data: List = None):
"""Initialize a speaker embedding layer if needen and define expected embedding channel size for defining
`in_channels` size of the connected layers.
This implementation yields 3 possible outcomes:
1. If `config.use_speaker_embedding` and `config.use_d_vector_file are False, do nothing.
2. If `config.use_d_vector_file` is True, set expected embedding channel size to `config.d_vector_dim` or 512.
3. If `config.use_speaker_embedding`, initialize a speaker embedding layer with channel size of
`config.d_vector_dim` or 512.
You can override this function for new models.0
Args:
config (Coqpit): Model configuration.
data (List, optional): Dataset items to infer number of speakers. Defaults to None.
"""
# init speaker manager
self.speaker_manager = get_speaker_manager(config, data=data)
# set number of speakers - if num_speakers is set in config, use it, otherwise use speaker_manager
if data is not None or self.speaker_manager.speaker_ids:
self.num_speakers = self.speaker_manager.num_speakers
else:
self.num_speakers = (config.num_speakers
if "num_speakers" in config
and config.num_speakers != 0 else
self.speaker_manager.num_speakers)
# set ultimate speaker embedding size
if config.use_speaker_embedding or config.use_d_vector_file:
self.embedded_speaker_dim = (
config.d_vector_dim if "d_vector_dim" in config
and config.d_vector_dim is not None else 512)
# init speaker embedding layer
if config.use_speaker_embedding and not config.use_d_vector_file:
self.speaker_embedding = nn.Embedding(self.num_speakers,
self.embedded_speaker_dim)
self.speaker_embedding.weight.data.normal_(0, 0.3)
def init_multispeaker(self, config: "Coqpit", data: list = None) -> None:
"""Initialize multi-speaker modules of a model. A model can be trained either with a speaker embedding layer
or with external `d_vectors` computed from a speaker encoder model.
If you need a different behaviour, override this function for your model.
Args:
config (Coqpit): Model configuration.
data (List, optional): Dataset items to infer number of speakers. Defaults to None.
"""
# init speaker manager
self.speaker_manager = get_speaker_manager(config, data=data)
self.num_speakers = self.speaker_manager.num_speakers
if config.use_d_vector_file:
self.external_d_vector_dim = config.d_vector_dim
else:
self.external_d_vector_dim = 0
# init speaker embedding layer
if config.use_speaker_embedding and not config.use_d_vector_file:
self.embedded_speaker_dim = self.c_in_channels
self.emb_g = nn.Embedding(self.num_speakers,
self.embedded_speaker_dim)
nn.init.uniform_(self.emb_g.weight, -0.1, 0.1)
def get_speaker_manager(config: Coqpit,
restore_path: str,
data: List,
out_path: str = None) -> SpeakerManager:
return get_speaker_manager(config, restore_path, data, out_path)