def forward(self, spect_predicted, log_durs_predicted, pitch_predicted,
spect_tgt, durs_tgt, dur_lens, pitch_tgt):
spect_tgt.requires_grad = False
spect_tgt = spect_tgt.transpose(1, 2) # (B, T, H)
dur_mask = mask_from_lens(dur_lens, max_len=durs_tgt.size(1))
log_durs_tgt = torch.log(durs_tgt.float() + 1)
loss_fn = F.mse_loss
dur_loss = loss_fn(log_durs_predicted, log_durs_tgt, reduction='none')
dur_loss = (dur_loss * dur_mask).sum() / dur_mask.sum()
dur_loss *= self.dur_predictor_loss_scale
ldiff = spect_tgt.size(1) - spect_predicted.size(1)
spect_predicted = F.pad(spect_predicted, (0, 0, 0, ldiff, 0, 0),
value=0.0)
mel_mask = spect_tgt.ne(0).float()
loss_fn = F.mse_loss
mel_loss = loss_fn(spect_predicted, spect_tgt, reduction='none')
mel_loss = (mel_loss * mel_mask).sum() / mel_mask.sum()
ldiff = pitch_tgt.size(1) - pitch_predicted.size(1)
pitch_predicted = F.pad(pitch_predicted, (0, ldiff, 0, 0), value=0.0)
pitch_loss = F.mse_loss(pitch_tgt, pitch_predicted, reduction='none')
pitch_loss = (pitch_loss * dur_mask).sum() / dur_mask.sum()
pitch_loss *= self.pitch_predictor_loss_scale
return mel_loss + pitch_loss + dur_loss, mel_loss, dur_loss, pitch_loss
def forward(self, pitch_predicted, pitch_tgt, len):
dur_mask = mask_from_lens(len, max_len=pitch_tgt.size(1))
ldiff = pitch_tgt.size(1) - pitch_predicted.size(1)
pitch_predicted = F.pad(pitch_predicted, (0, ldiff, 0, 0), value=0.0)
pitch_loss = F.mse_loss(pitch_tgt, pitch_predicted, reduction='none')
pitch_loss = (pitch_loss * dur_mask).sum() / dur_mask.sum()
pitch_loss *= self.loss_scale
return pitch_loss
def forward(self, log_durs_predicted, durs_tgt, len):
dur_mask = mask_from_lens(len, max_len=durs_tgt.size(1))
log_durs_tgt = torch.log(durs_tgt.float() + 1)
loss_fn = F.mse_loss
dur_loss = loss_fn(log_durs_predicted, log_durs_tgt, reduction='none')
dur_loss = (dur_loss * dur_mask).sum() / dur_mask.sum()
dur_loss *= self.loss_scale
return dur_loss
def forward(self, log_durs_predicted, pitch_predicted, durs_tgt, dur_lens,
pitch_tgt):
dur_mask = mask_from_lens(dur_lens, max_len=durs_tgt.size(1))
log_durs_tgt = torch.log(durs_tgt.float() + 1)
loss_fn = F.mse_loss
dur_loss = loss_fn(log_durs_predicted, log_durs_tgt, reduction='none')
dur_loss = (dur_loss * dur_mask).sum() / dur_mask.sum()
dur_loss *= self.dur_predictor_loss_scale
ldiff = pitch_tgt.size(1) - pitch_predicted.size(1)
pitch_predicted = F.pad(pitch_predicted, (0, ldiff, 0, 0), value=0.0)
pitch_loss = F.mse_loss(pitch_tgt, pitch_predicted, reduction='none')
pitch_loss = (pitch_loss * dur_mask).sum() / dur_mask.sum()
pitch_loss *= self.pitch_predictor_loss_scale
return pitch_loss + dur_loss, dur_loss, pitch_loss