Medix-AI/coqui-tts
Medix-AI
/
coqui-tts
mirror of https://github.com/coqui-ai/TTS.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

differential spectral loss

This commit is contained in:
erogol 2020-09-24 12:38:11 +02:00
parent e0d4b88877
commit 6f0654f9a8
1 changed files with 13 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
TTS/tts/layers/losses.py
View File

@ -200,7 +200,7 @@ class TacotronLoss(torch.nn.Module):
self.criterion = nn.L1Loss() if c.model in ["Tacotron" self.criterion = nn.L1Loss() if c.model in ["Tacotron"
] else nn.MSELoss() ] else nn.MSELoss()
# differential spectral loss # differential spectral loss
if c.diff_spec_loss_alpha > 0: if c.diff_spec_alpha > 0:
self.criterion_diff_spec = DifferentailSpectralLoss(loss_func=self.criterion) self.criterion_diff_spec = DifferentailSpectralLoss(loss_func=self.criterion)
# guided attention loss # guided attention loss
if c.ga_alpha > 0: if c.ga_alpha > 0:
@ -254,11 +254,17 @@ class TacotronLoss(torch.nn.Module):
torch.flip(decoder_b_output, dims=(1, )), mel_input, torch.flip(decoder_b_output, dims=(1, )), mel_input,
output_lens) output_lens)
else: else:
<<<<<<< HEAD
decoder_b_loss = self.criterion( decoder_b_loss = self.criterion(
torch.flip(decoder_b_output, dims=(1, )), mel_input) torch.flip(decoder_b_output, dims=(1, )), mel_input)
decoder_c_loss = torch.nn.functional.l1_loss( decoder_c_loss = torch.nn.functional.l1_loss(
torch.flip(decoder_b_output, dims=(1, )), decoder_output) torch.flip(decoder_b_output, dims=(1, )), decoder_output)
loss += decoder_b_loss + decoder_c_loss loss += decoder_b_loss + decoder_c_loss
=======
decoder_b_loss = self.criterion(torch.flip(decoder_b_output, dims=(1, )), mel_input)
decoder_c_loss = torch.nn.functional.l1_loss(torch.flip(decoder_b_output, dims=(1, )), decoder_output)
loss += self.decoder_alpha * (decoder_b_loss + decoder_c_loss)
>>>>>>> differential spectral loss
return_dict['decoder_b_loss'] = decoder_b_loss return_dict['decoder_b_loss'] = decoder_b_loss
return_dict['decoder_c_loss'] = decoder_c_loss return_dict['decoder_c_loss'] = decoder_c_loss
@ -267,9 +273,14 @@ class TacotronLoss(torch.nn.Module):
decoder_b_loss = self.criterion(decoder_b_output, mel_input, decoder_b_loss = self.criterion(decoder_b_output, mel_input,
output_lens) output_lens)
# decoder_c_loss = torch.nn.functional.l1_loss(decoder_b_output, decoder_output) # decoder_c_loss = torch.nn.functional.l1_loss(decoder_b_output, decoder_output)
<<<<<<< HEAD
attention_c_loss = torch.nn.functional.l1_loss( attention_c_loss = torch.nn.functional.l1_loss(
alignments, alignments_backwards) alignments, alignments_backwards)
loss += decoder_b_loss + attention_c_loss loss += decoder_b_loss + attention_c_loss
=======
attention_c_loss = torch.nn.functional.l1_loss(alignments, alignments_backwards)
loss += self.decoder_alpha * (decoder_b_loss + attention_c_loss)
>>>>>>> differential spectral loss
return_dict['decoder_coarse_loss'] = decoder_b_loss return_dict['decoder_coarse_loss'] = decoder_b_loss
return_dict['decoder_ddc_loss'] = attention_c_loss return_dict['decoder_ddc_loss'] = attention_c_loss
@ -280,7 +291,7 @@ class TacotronLoss(torch.nn.Module):
return_dict['ga_loss'] = ga_loss * self.ga_alpha return_dict['ga_loss'] = ga_loss * self.ga_alpha
# differential spectral loss # differential spectral loss
if self.config.diff_spec_loss_alpha > 0: if self.config.diff_spec_alpha > 0:
diff_spec_loss = self.criterion_diff_spec(postnet_output, mel_input, output_lens) diff_spec_loss = self.criterion_diff_spec(postnet_output, mel_input, output_lens)
loss += diff_spec_loss * self.diff_spec_alpha loss += diff_spec_loss * self.diff_spec_alpha
return_dict['diff_spec_loss'] = diff_spec_loss return_dict['diff_spec_loss'] = diff_spec_loss