def main():
"""Run `tts` model training directly by a `config.json` file."""
# init trainer args
train_args = TrainVocoderArgs()
parser = train_args.init_argparse(arg_prefix="")
# override trainer args from comman-line args
args, config_overrides = parser.parse_known_args()
train_args.parse_args(args)
# load config.json and register
if args.config_path or args.continue_path:
if args.config_path:
# init from a file
config = load_config(args.config_path)
if len(config_overrides) > 0:
config.parse_known_args(config_overrides, relaxed_parser=True)
elif args.continue_path:
# continue from a prev experiment
config = load_config(
os.path.join(args.continue_path, "config.json"))
if len(config_overrides) > 0:
config.parse_known_args(config_overrides, relaxed_parser=True)
else:
# init from console args
from TTS.config.shared_configs import BaseTrainingConfig # pylint: disable=import-outside-toplevel
config_base = BaseTrainingConfig()
config_base.parse_known_args(config_overrides)
config = register_config(config_base.model)()
# load training samples
if "feature_path" in config and config.feature_path:
# load pre-computed features
print(f" > Loading features from: {config.feature_path}")
eval_samples, train_samples = load_wav_feat_data(
config.data_path, config.feature_path, config.eval_split_size)
else:
# load data raw wav files
eval_samples, train_samples = load_wav_data(config.data_path,
config.eval_split_size)
# setup audio processor
ap = AudioProcessor(**config.audio)
# init the model from config
model = setup_model(config)
# init the trainer and 🚀
trainer = Trainer(
train_args,
config,
config.output_path,
model=model,
train_samples=train_samples,
eval_samples=eval_samples,
training_assets={"audio_processor": ap},
parse_command_line_args=False,
)
trainer.fit()
def main():
"""Run `tts` model training directly by a `config.json` file."""
# init trainer args
train_args = TrainTTSArgs()
parser = train_args.init_argparse(arg_prefix="")
# override trainer args from comman-line args
args, config_overrides = parser.parse_known_args()
train_args.parse_args(args)
# load config.json and register
if args.config_path or args.continue_path:
if args.config_path:
# init from a file
config = load_config(args.config_path)
if len(config_overrides) > 0:
config.parse_known_args(config_overrides, relaxed_parser=True)
elif args.continue_path:
# continue from a prev experiment
config = load_config(os.path.join(args.continue_path, "config.json"))
if len(config_overrides) > 0:
config.parse_known_args(config_overrides, relaxed_parser=True)
else:
# init from console args
from TTS.config.shared_configs import BaseTrainingConfig # pylint: disable=import-outside-toplevel
config_base = BaseTrainingConfig()
config_base.parse_known_args(config_overrides)
config = register_config(config_base.model)()
# load training samples
train_samples, eval_samples = load_tts_samples(
config.datasets,
eval_split=True,
eval_split_max_size=config.eval_split_max_size,
eval_split_size=config.eval_split_size,
)
# init the model from config
model = setup_model(config, train_samples + eval_samples)
# init the trainer and 🚀
trainer = Trainer(
train_args,
model.config,
config.output_path,
model=model,
train_samples=train_samples,
eval_samples=eval_samples,
parse_command_line_args=False,
)
trainer.fit()
def main():
"""Run `tts` model training directly by a `config.json` file."""
# init trainer args
train_args = TrainingArgs()
parser = train_args.init_argparse(arg_prefix="")
# override trainer args from comman-line args
args, config_overrides = parser.parse_known_args()
train_args.parse_args(args)
# load config.json and register
if args.config_path or args.continue_path:
if args.config_path:
# init from a file
config = load_config(args.config_path)
if len(config_overrides) > 0:
config.parse_known_args(config_overrides, relaxed_parser=True)
elif args.continue_path:
# continue from a prev experiment
config = load_config(
os.path.join(args.continue_path, "config.json"))
if len(config_overrides) > 0:
config.parse_known_args(config_overrides, relaxed_parser=True)
else:
# init from console args
from TTS.config.shared_configs import BaseTrainingConfig # pylint: disable=import-outside-toplevel
config_base = BaseTrainingConfig()
config_base.parse_known_args(config_overrides)
config = register_config(config_base.model)()
# load training samples
train_samples, eval_samples = load_tts_samples(config.datasets,
eval_split=True)
# setup audio processor
ap = AudioProcessor(**config.audio)
# init speaker manager
if check_config_and_model_args(config, "use_speaker_embedding", True):
speaker_manager = SpeakerManager(data_items=train_samples +
eval_samples)
if hasattr(config, "model_args"):
config.model_args.num_speakers = speaker_manager.num_speakers
else:
config.num_speakers = speaker_manager.num_speakers
elif check_config_and_model_args(config, "use_d_vector_file", True):
if check_config_and_model_args(config, "use_speaker_encoder_as_loss",
True):
speaker_manager = SpeakerManager(
d_vectors_file_path=config.model_args.d_vector_file,
encoder_model_path=config.model_args.
speaker_encoder_model_path,
encoder_config_path=config.model_args.
speaker_encoder_config_path,
use_cuda=torch.cuda.is_available(),
)
else:
speaker_manager = SpeakerManager(
d_vectors_file_path=get_from_config_or_model_args(
config, "d_vector_file"))
config.num_speakers = speaker_manager.num_speakers
if hasattr(config, "model_args"):
config.model_args.num_speakers = speaker_manager.num_speakers
else:
speaker_manager = None
if check_config_and_model_args(config, "use_language_embedding", True):
language_manager = LanguageManager(config=config)
if hasattr(config, "model_args"):
config.model_args.num_languages = language_manager.num_languages
else:
config.num_languages = language_manager.num_languages
else:
language_manager = None
# init the model from config
model = setup_model(config, speaker_manager, language_manager)
# init the trainer and 🚀
trainer = Trainer(
train_args,
config,
config.output_path,
model=model,
train_samples=train_samples,
eval_samples=eval_samples,
training_assets={"audio_processor": ap},
parse_command_line_args=False,
)
trainer.fit()
def process_args(args, config=None):
"""Process parsed comand line arguments and initialize the config if not provided.
Args:
args (argparse.Namespace or dict like): Parsed input arguments.
config (Coqpit): Model config. If none, it is generated from `args`. Defaults to None.
Returns:
c (TTS.utils.io.AttrDict): Config paramaters.
out_path (str): Path to save models and logging.
audio_path (str): Path to save generated test audios.
c_logger (TTS.utils.console_logger.ConsoleLogger): Class that does
logging to the console.
dashboard_logger (WandbLogger or TensorboardLogger): Class that does the dashboard Logging
TODO:
- Interactive config definition.
"""
if isinstance(args, tuple):
args, coqpit_overrides = args
if args.continue_path:
# continue a previous training from its output folder
experiment_path = args.continue_path
args.config_path = os.path.join(args.continue_path, "config.json")
args.restore_path, best_model = get_last_checkpoint(args.continue_path)
if not args.best_path:
args.best_path = best_model
# init config if not already defined
if config is None:
if args.config_path:
# init from a file
config = load_config(args.config_path)
else:
# init from console args
from TTS.config.shared_configs import BaseTrainingConfig # pylint: disable=import-outside-toplevel
config_base = BaseTrainingConfig()
config_base.parse_known_args(coqpit_overrides)
config = register_config(config_base.model)()
# override values from command-line args
config.parse_known_args(coqpit_overrides, relaxed_parser=True)
experiment_path = args.continue_path
if not experiment_path:
experiment_path = get_experiment_folder_path(config.output_path,
config.run_name)
audio_path = os.path.join(experiment_path, "test_audios")
config.output_log_path = experiment_path
# setup rank 0 process in distributed training
dashboard_logger = None
if args.rank == 0:
new_fields = {}
if args.restore_path:
new_fields["restore_path"] = args.restore_path
new_fields["github_branch"] = get_git_branch()
# if model characters are not set in the config file
# save the default set to the config file for future
# compatibility.
if config.has("characters") and config.characters is None:
used_characters = parse_symbols()
new_fields["characters"] = used_characters
copy_model_files(config, experiment_path, new_fields)
dashboard_logger = logger_factory(config, experiment_path)
c_logger = ConsoleLogger()
return config, experiment_path, audio_path, c_logger, dashboard_logger