def __init__(self, data_utterance_paths, hparams):
"""Data loader for the PPG->Mel task.
Args:
data_utterance_paths: A text file containing a list of file paths.
hparams: The hyper-parameters.
"""
self.data_utterance_paths = load_filepaths(data_utterance_paths)
self.max_wav_value = hparams.max_wav_value
self.sampling_rate = hparams.sampling_rate
self.is_full_ppg = hparams.is_full_ppg
self.is_append_f0 = hparams.is_append_f0
self.is_cache_feats = hparams.is_cache_feats
self.load_feats_from_disk = hparams.load_feats_from_disk
self.feats_cache_path = hparams.feats_cache_path
self.ppg_subsampling_factor = hparams.ppg_subsampling_factor
self.ppg_deps = DependenciesPPG()
self.encoder_model_fpath = hparams.encoder_model_fpath
if self.is_cache_feats and self.load_feats_from_disk:
raise ValueError('If you are loading feats from the disk, do not '
'rewrite them back!')
self.stft = layers.TacotronSTFT(
hparams.filter_length, hparams.hop_length, hparams.win_length,
hparams.n_acoustic_feat_dims, hparams.sampling_rate,
hparams.mel_fmin, hparams.mel_fmax)
random.seed(hparams.seed)
random.shuffle(self.data_utterance_paths)
self.ppg_sequences = []
self.acoustic_sequences = []
self.dvec_sequences = []
if self.load_feats_from_disk:
print('Loading data from %s.' % self.feats_cache_path)
with open(self.feats_cache_path, 'rb') as f:
data = pickle.load(f)
self.ppg_sequences = data[0]
self.acoustic_sequences = data[1]
self.dvec_sequences = data[2]
else:
for utterance_path in self.data_utterance_paths:
ppg_feat_pair_spkr = self.extract_utterance_feats_spkr(utterance_path, self.is_full_ppg)
self.ppg_sequences.append(ppg_feat_pair_spkr[0].astype(np.float32))
self.acoustic_sequences.append(ppg_feat_pair_spkr[1])
self.dvec_sequences.append(ppg_feat_pair_spkr[2])
if self.is_cache_feats:
print('Caching data to %s.' % self.feats_cache_path)
with open(self.feats_cache_path, 'wb') as f:
pickle.dump([self.ppg_sequences, self.acoustic_sequences, self.dvec_sequences], f)
def __init__(self, data_utterance_paths, cache_path, hparams, bs, loop):
"""Data loader for the PPG->Mel task.
Args:
data_utterance_paths: A text file containing a list of file paths.
hparams: The hyper-parameters.
"""
self.data_utterance_paths = load_filepaths(data_utterance_paths)
self.max_wav_value = 32768.0
self.sampling_rate = hparams.sampling_rate
self.is_full_ppg = True
self.is_append_f0 = False
self.is_cache_feats = True
self.feats_cache_path = cache_path
self.ppg_subsampling_factor = 1
self.ppg_deps = DependenciesPPG()
# 20 data = n(4) * b(5)
self.n = int(bs) - 1
self.b = 5
self.l = int(loop) - 1
self.stft = layers.TacotronSTFT(
hparams.filter_length, hparams.hop_length, hparams.win_length,
hparams.n_acoustic_feat_dims, self.sampling_rate,
hparams.mel_fmin, hparams.mel_fmax)
random.seed(hparams.seed)
random.shuffle(self.data_utterance_paths)
if self.n > 0:
with open(self.feats_cache_path, 'rb') as f:
data = pickle.load(f)
self.ppg_sequences = data[0]
self.acoustic_sequences = data[1]
else:
self.ppg_sequences = []
self.acoustic_sequences = []
for utterance_path in self.data_utterance_paths[self.n * self.b + self.l * 20 : (self.n+1) * self.b + self.l * 20]:
ppg_feat_pair = self.extract_utterance_feats(utterance_path,
self.is_full_ppg)
self.ppg_sequences.append(ppg_feat_pair[0].astype(
np.float32))
self.acoustic_sequences.append(ppg_feat_pair[1])
if self.is_cache_feats:
print('Caching data to %s.' % self.feats_cache_path)
with open(self.feats_cache_path, 'wb+') as f:
pickle.dump([self.ppg_sequences, self.acoustic_sequences], f)