def _get_model(n_speakers=1, speaker_embed_dim=None):
model = build_deepvoice3(
n_vocab=n_vocab,
embed_dim=256,
mel_dim=num_mels,
linear_dim=num_freq,
r=outputs_per_step,
padding_idx=padding_idx,
n_speakers=n_speakers,
speaker_embed_dim=speaker_embed_dim,
)
return model
def build_model():
model = build_deepvoice3(n_vocab=_frontend.n_vocab,
embed_dim=hparams.text_embed_dim,
mel_dim=hparams.num_mels,
linear_dim=hparams.fft_size // 2 + 1,
r=hparams.outputs_per_step,
padding_idx=hparams.padding_idx,
dropout=hparams.dropout,
kernel_size=hparams.kernel_size,
encoder_channels=hparams.encoder_channels,
decoder_channels=hparams.decoder_channels,
converter_channels=hparams.converter_channels,
use_memory_mask=hparams.use_memory_mask,
trainable_positional_encodings=hparams.trainable_positional_encodings
)
return model
def _get_model(n_speakers=1, speaker_embed_dim=None):
model = build_deepvoice3(
n_vocab=n_vocab,
embed_dim=256,
mel_dim=num_mels,
linear_dim=num_freq,
r=outputs_per_step,
padding_idx=padding_idx,
n_speakers=n_speakers,
speaker_embed_dim=speaker_embed_dim,
dropout=1 - 0.95,
kernel_size=5,
encoder_channels=128,
decoder_channels=256,
converter_channels=256,
)
return model
max_decoder_steps = int(args["--max-decoder-steps"])
file_name_suffix = args["--file-name-suffix"]
# Override hyper parameters
hparams.parse(args["--hparams"])
assert hparams.name == "deepvoice3"
_frontend = getattr(frontend, hparams.frontend)
# Model
model = build_deepvoice3(n_vocab=_frontend.n_vocab,
embed_dim=256,
mel_dim=hparams.num_mels,
linear_dim=hparams.num_freq,
r=hparams.outputs_per_step,
padding_idx=hparams.padding_idx,
dropout=hparams.dropout,
kernel_size=hparams.kernel_size,
encoder_channels=hparams.encoder_channels,
decoder_channels=hparams.decoder_channels,
converter_channels=hparams.converter_channels,
)
checkpoint = torch.load(checkpoint_path)
model.load_state_dict(checkpoint["state_dict"])
model.decoder.max_decoder_steps = max_decoder_steps
model.make_generation_fast_()
os.makedirs(dst_dir, exist_ok=True)
with open(text_list_file_path, "rb") as f:
lines = f.readlines()