Support for pytorch v0.4

[r9y9]

No description provided.

[engiecat]

There seems to be an error caused by new changes.
When I trigger synthesis, the following error occurs

Traceback (most recent call last):
  File "synthesis.py", line 144, in <module>
    model, text, p=replace_pronunciation_prob, speaker_id=speaker_id, fast=True)
  File "synthesis.py", line 67, in tts
    sequence, text_positions=text_positions, speaker_ids=speaker_ids)
  File "H:\envs\pytorch\lib\site-packages\torch\nn\modules\module.py", line 491, in __call__
    result = self.forward(*input, **kwargs)
  File "H:\Tensorflow_Study\git\dv3_torch04\deepvoice3_pytorch\__init__.py", line 79, in forward
    text_positions, frame_positions, input_lengths)
  File "H:\envs\pytorch\lib\site-packages\torch\nn\modules\module.py", line 491, in __call__
    result = self.forward(*input, **kwargs)
  File "H:\Tensorflow_Study\git\dv3_torch04\deepvoice3_pytorch\__init__.py", line 124, in forward
    speaker_embed=speaker_embed, lengths=input_lengths)
  File "H:\envs\pytorch\lib\site-packages\torch\nn\modules\module.py", line 491, in __call__
    result = self.forward(*input, **kwargs)
  File "H:\Tensorflow_Study\git\dv3_torch04\deepvoice3_pytorch\deepvoice3.py", line 284, in forward
    outputs = self.incremental_forward(encoder_out, text_positions, speaker_embed)
  File "H:\Tensorflow_Study\git\dv3_torch04\deepvoice3_pytorch\deepvoice3.py", line 420, in incremental_forward
    x = f.incremental_forward(x, speaker_embed)
  File "H:\Tensorflow_Study\git\dv3_torch04\deepvoice3_pytorch\modules.py", line 143, in incremental_forward
    return self._forward(x, speaker_embed, True)
  File "H:\Tensorflow_Study\git\dv3_torch04\deepvoice3_pytorch\modules.py", line 150, in _forward
    x = self.conv.incremental_forward(x)
  File "H:\Tensorflow_Study\git\dv3_torch04\deepvoice3_pytorch\conv.py", line 43, in incremental_forward
    input = torch.Tensor(self.input_buffer)
TypeError: expected torch.FloatTensor (got torch.cuda.FloatTensor)

For now, I had created a quick fix by changing line43 of conv.py to
input = torch.Tensor(self.input_buffer.cpu()).cuda()
but I think that this will cause performance drop.

[r9y9]

Thanks for the report. I will look into it.

[r9y9]

Fixed in 7b6f82f

[r9y9]

Turn out that after fixing for pytorch v0.4, I cannot get reasonable speech quality. Not sure what is the problem yet, I should look into it closely..,

[r9y9]

For the record, 7907592 with deepvoice3_ljspeech preset works

[engiecat]

just confirmed that at least synthesis.py works equally between pytorch 0.3 version and current version.
(Same checkpoint file, same text -> same output audio)

[r9y9]

Thanks for letting me know. Yes I mean training deepvoice3 doesn't work anymore.

[r9y9]

As for the issue I previously mentioned was actually a problem of eval_model in train.py. There's no issue for synthesis as you said. Fixed by ec4cf2a, not sure what's actually the problem is. It's hard to debug so for now I changed the args of tts function as same as synthesis.py; i.e., fast=True.