def forward(self, batch_wave, lengths, phone, len_phone, label_smooth=0., threshold=0.95):
device = batch_wave.device
phone_paddings = phone.eq(0).float()
encoder_outputs, encoder_output_lengths = self.splayer(batch_wave, lengths)
encoder_outputs, encoder_output_lengths = self.encoder(encoder_outputs, encoder_output_lengths)
ctc_logits = self.ctc_fc(encoder_outputs)
len_logits_ctc = encoder_output_lengths
alphas = self.assigner(encoder_outputs, encoder_output_lengths)
# sum
_num = alphas.sum(-1)
# scaling
num = len_phone.float()
num_noise = num + 0.9 * torch.rand(alphas.size(0)).to(device) - 0.45
alphas *= (num_noise / _num)[:, None].repeat(1, alphas.size(1))
cif_outputs = self.cif(encoder_outputs, alphas, threshold=threshold)
logits_IPA = self.phone_fc(cif_outputs)
ctc_loss = cal_ctc_loss(ctc_logits, len_logits_ctc, phone, len_phone)
qua_loss = cal_qua_loss(_num, num)
ce_phone_loss = cal_ce_loss(logits_IPA, phone, phone_paddings, label_smooth)
return ctc_loss, qua_loss, ce_phone_loss
def forward(self, batch_wave, lengths, target_ids, target_labels=None, target_paddings=None, label_smooth=0., threshold=0.95):
device = batch_wave.device
target_lengths = torch.sum(1-target_paddings, dim=-1).long()
encoder_outputs, encoder_output_lengths = self.splayer(batch_wave, lengths)
encoder_outputs, encoder_output_lengths = self.encoder(encoder_outputs, encoder_output_lengths)
ctc_logits = self.ctc_fc(encoder_outputs)
len_logits_ctc = encoder_output_lengths
alphas = self.assigner(encoder_outputs, encoder_output_lengths)
# sum
_num = alphas.sum(-1)
# scaling
num = target_lengths.float()
num_noise = num + 0.9 * torch.rand(alphas.size(0)).to(device) - 0.45
alphas *= (num_noise / _num)[:, None].repeat(1, alphas.size(1))
cif_outputs = self.cif(encoder_outputs, alphas, threshold=threshold)
logits = self.decoder(cif_outputs, target_ids, target_lengths)
ctc_loss = cal_ctc_loss(ctc_logits, len_logits_ctc, target_labels, target_lengths)
qua_loss = cal_qua_loss(_num, num)
ce_loss = cal_ce_loss(logits, target_labels, target_paddings, label_smooth)
return ctc_loss, qua_loss, ce_loss
def forward(self, batch_wave, lengths, target_ids, target_labels=None, target_paddings=None, label_smooth=0.):
target_lengths = torch.sum(1-target_paddings, dim=-1).long()
logits = self.get_logits(batch_wave, lengths,
target_ids, target_lengths)
loss = cal_ce_loss(logits, target_labels, target_paddings, label_smooth)
return loss
def forward(self, batch_wave, lengths, target_ids, target_labels=None, target_paddings=None, label_smooth=0.):
target_lengths = torch.sum(1-target_paddings, dim=-1).long()
ctc_logits, len_logits_ctc, ce_logits = self.get_logits(
batch_wave, lengths, target_ids, target_lengths)
ctc_loss = cal_ctc_loss(ctc_logits, len_logits_ctc, target_labels, target_lengths-1) # the target of ctc counts without blk
ce_loss = cal_ce_loss(ce_logits, target_labels, target_paddings, label_smooth)
return ctc_loss, ce_loss
def forward(self,
tokens_input,
len_input,
target_input,
target_output,
target_paddings,
label_smooth=0.):
len_targets = torch.sum(1 - target_paddings, dim=-1).long()
logits = self.get_logits(tokens_input, len_input, target_input,
len_targets)
loss = cal_ce_loss(logits, target_output, target_paddings,
label_smooth)
return loss