from dv3.train import TextDataSource, MelSpecDataSource
from nnmnkwii.datasets import FileSourceDataset
from tqdm import trange
from dv3.deepvoice3_pytorch import frontend
if __name__ == "__main__":
args = docopt(__doc__)
data_root = args["<data_root>"]
preset = args["--preset"]
# Load preset if specified
if preset is not None:
with open(preset) as f:
hparams.parse_json(f.read())
# Override hyper parameters
hparams.parse(args["--hparams"])
assert hparams.name == "deepvoice3"
train._frontend = getattr(frontend, hparams.frontend)
# Code below
X = FileSourceDataset(TextDataSource(data_root))
Mel = FileSourceDataset(MelSpecDataSource(data_root))
in_sizes = []
out_sizes = []
for i in trange(len(X)):
x, m = X[i], Mel[i]
if X.file_data_source.multi_speaker:
x = x[0]
in_sizes.append(x.shape[0])
train_postnet = args["--train-postnet-only"]
# train both if not specified
if not train_seq2seq and not train_postnet:
print("Training whole model")
train_seq2seq, train_postnet = True, True
if train_seq2seq:
print("Training seq2seq model")
elif train_postnet:
print("Training postnet model")
else:
assert False, "must be specified wrong args"
# Override hyper parameters
# For Speaker Adaptation
hparams.parse(
'builder=deepvoice3_multispeaker,preset=deepvoice3_speaker_adaptation_vctk'
)
print(hparams_debug_string())
assert hparams.name == "deepvoice3"
# Presets
if hparams.preset is not None and hparams.preset != "":
preset = hparams.presets[hparams.preset]
import json
hparams.parse_json(json.dumps(preset))
print("Override hyper parameters with preset \"{}\": {}".format(
hparams.preset, json.dumps(preset, indent=4)))
_frontend = getattr(frontend, hparams.frontend)
os.makedirs(checkpoint_dir, exist_ok=True)
