def parse_batch(self, batch):
audio, attention_contexts, encoder_outputs, text_lengths, durations = batch
audio = to_gpu(audio).float()
attention_contexts = to_gpu(attention_contexts).float()
encoder_outputs = to_gpu(encoder_outputs).float()
text_lengths = to_gpu(text_lengths).long()
durations = to_gpu(durations).long()
return (audio, attention_contexts, encoder_outputs, text_lengths,
durations)
def calculate_global_mean(data_loader, global_mean_npy, hparams):
if global_mean_npy and os.path.exists(global_mean_npy):
global_mean = np.load(global_mean_npy)
return to_gpu(torch.tensor(global_mean).half()) if hparams.fp16_run else to_gpu(torch.tensor(global_mean).float())
sums = []
frames = []
print('calculating global mean...')
for i, batch in tqdm(enumerate(data_loader), total=len(data_loader), smoothing=0.001):
# padded values are 0.
sums.append(batch['gt_mel'].double().sum(dim=(0, 2)))
frames.append(batch['mel_lengths'].double().sum())
if i > 100:
break
global_mean = sum(sums) / sum(frames)
global_mean = to_gpu(global_mean.half()) if hparams.fp16_run else to_gpu(global_mean.float())
if global_mean_npy:
np.save(global_mean_npy, global_mean.cpu().numpy())
return global_mean
def calculate_global_mean(data_loader, global_mean_npy, hparams):
if global_mean_npy and os.path.exists(global_mean_npy):
global_mean = np.load(global_mean_npy)
return to_gpu(torch.tensor(global_mean).half()) if hparams.fp16_run else to_gpu(torch.tensor(global_mean).float())
sums = []
frames = []
print('calculating global mean...')
for i, batch in tqdm(enumerate(data_loader), total=len(data_loader), smoothing=0.001):
text_padded, input_lengths, mel_padded, gate_padded,\
output_lengths, speaker_ids, torchmoji_hidden, preserve_decoder_states = batch
# padded values are 0.
sums.append(mel_padded.double().sum(dim=(0, 2)))
frames.append(output_lengths.double().sum())
global_mean = sum(sums) / sum(frames)
global_mean = to_gpu(global_mean.half()) if hparams.fp16_run else to_gpu(global_mean.float())
if global_mean_npy:
np.save(global_mean_npy, global_mean.cpu().numpy())
return global_mean
def parse_batch(self, batch):
text_padded, text_lengths, mel_padded, gate_padded, \
output_lengths, speaker_ids, torchmoji_hidden, preserve_decoder_states = batch
text_padded = to_gpu(text_padded).long()
text_lengths = to_gpu(text_lengths).long()
output_lengths = to_gpu(output_lengths).long()
speaker_ids = to_gpu(speaker_ids.data).long()
mel_padded = to_gpu(mel_padded).float()
max_len = torch.max(text_lengths.data).item() # used by loss func
gate_padded = to_gpu(gate_padded).float() # used by loss func
if torchmoji_hidden is not None:
torchmoji_hidden = to_gpu(torchmoji_hidden).float()
if preserve_decoder_states is not None:
preserve_decoder_states = to_gpu(preserve_decoder_states).float()
return (
(text_padded, text_lengths, mel_padded, max_len, output_lengths, speaker_ids, torchmoji_hidden, preserve_decoder_states),
(mel_padded, gate_padded, output_lengths, text_lengths))
def parse_batch(self, batch):
text_padded, mel_padded, speaker_ids, text_lengths, output_lengths,\
alignments, torchmoji_hidden, perc_loudness, f0, energy, sylps,\
voiced_mask, char_f0, char_voiced, char_energy = batch
text_padded = to_gpu(text_padded).long()
mel_padded = to_gpu(mel_padded).float()
speaker_ids = to_gpu(speaker_ids.data).long()
text_lengths = to_gpu(text_lengths).long()
output_lengths = to_gpu(output_lengths).long()
alignments = to_gpu(alignments).float()
if torchmoji_hidden is not None:
torchmoji_hidden = to_gpu(torchmoji_hidden).float()
perc_loudness = to_gpu(perc_loudness).float()
f0 = to_gpu(f0).float()
energy = to_gpu(energy).float()
sylps = to_gpu(sylps).float()
voiced_mask = to_gpu(voiced_mask).bool()
char_f0 = to_gpu(char_f0).float()
char_voiced = to_gpu(char_voiced).float()
char_energy = to_gpu(char_energy).float()
return ((text_padded, mel_padded, speaker_ids, text_lengths,
output_lengths, alignments, torchmoji_hidden, perc_loudness,
f0, energy, sylps, voiced_mask, char_f0, char_voiced,
char_energy), (mel_padded, text_lengths, output_lengths,
perc_loudness, f0, energy, sylps, voiced_mask,
char_f0, char_voiced, char_energy))