def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None):
if isinstance(cfg, dict):
cfg = OmegaConf.create(cfg)
super().__init__(cfg=cfg, trainer=trainer)
schema = OmegaConf.structured(FastSpeech2Config)
# ModelPT ensures that cfg is a DictConfig, but do this second check in case ModelPT changes
if isinstance(cfg, dict):
cfg = OmegaConf.create(cfg)
elif not isinstance(cfg, DictConfig):
raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig")
# Ensure passed cfg is compliant with schema
OmegaConf.merge(cfg, schema)
self.pitch = cfg.add_pitch_predictor
self.energy = cfg.add_energy_predictor
self.duration_coeff = cfg.duration_coeff
self.audio_to_melspec_preprocessor = instantiate(self._cfg.preprocessor)
self.encoder = instantiate(self._cfg.encoder)
self.mel_decoder = instantiate(self._cfg.decoder)
self.variance_adapter = instantiate(self._cfg.variance_adaptor)
self.loss = L2MelLoss()
self.mseloss = torch.nn.MSELoss()
self.durationloss = DurationLoss()
self.log_train_images = False
# Parser and mappings are used for inference only.
self.parser = parsers.make_parser(name='en')
with open(cfg.mappings_filepath, 'r') as f:
mappings = json.load(f)
self.word2phones = mappings['word2phones']
self.phone2idx = mappings['phone2idx']
def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None):
if isinstance(cfg, dict):
cfg = OmegaConf.create(cfg)
super().__init__(cfg=cfg, trainer=trainer)
self.audio_to_melspec_precessor = instantiate(cfg.preprocessor)
self.encoder = instantiate(cfg.encoder)
self.variance_adapter = instantiate(cfg.variance_adaptor)
self.generator = instantiate(cfg.generator)
self.multiperioddisc = MultiPeriodDiscriminator()
self.multiscaledisc = MultiScaleDiscriminator()
self.melspec_fn = instantiate(cfg.preprocessor,
highfreq=None,
use_grads=True)
self.mel_val_loss = L1MelLoss()
self.durationloss = DurationLoss()
self.feat_matching_loss = FeatureMatchingLoss()
self.disc_loss = DiscriminatorLoss()
self.gen_loss = GeneratorLoss()
self.mseloss = torch.nn.MSELoss()
self.energy = cfg.add_energy_predictor
self.pitch = cfg.add_pitch_predictor
self.mel_loss_coeff = cfg.mel_loss_coeff
self.pitch_loss_coeff = cfg.pitch_loss_coeff
self.energy_loss_coeff = cfg.energy_loss_coeff
self.splice_length = cfg.splice_length
self.use_energy_pred = False
self.use_pitch_pred = False
self.log_train_images = False
self.logged_real_samples = False
self._tb_logger = None
self.sample_rate = cfg.sample_rate
self.hop_size = cfg.hop_size
# Parser and mappings are used for inference only.
self.parser = parsers.make_parser(name='en')
if 'mappings_filepath' in cfg:
mappings_filepath = cfg.get('mappings_filepath')
else:
logging.error(
"ERROR: You must specify a mappings.json file in the config file under model.mappings_filepath."
)
mappings_filepath = self.register_artifact('mappings_filepath',
mappings_filepath)
with open(mappings_filepath, 'r') as f:
mappings = json.load(f)
self.word2phones = mappings['word2phones']
self.phone2idx = mappings['phone2idx']