mrq/DL-Art-School
1
3
Fork 5
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Bug fixes for tacotron dataset on mozilla cv

Browse Source 
- Support a max mel length (mozilla cv has some tracks that are basically unbounded..)
- Don't fail on low sample rates (mozilla cv has some of those)
This commit is contained in:
James Betker 2021-08-11 16:17:03 -06:00
parent 98f37241cf
commit f04a7bdf63
1 changed files with 10 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
codes/data/audio/nv_tacotron_dataset.py
View File

@ -40,6 +40,7 @@ class TextMelLoader(torch.utils.data.Dataset):
self.audiopaths_and_text = fetcher_fn(hparams['path'])
self.text_cleaners = hparams.text_cleaners
self.max_wav_value = hparams.max_wav_value
self.max_mel_len = opt_get(hparams, ['max_mel_length'], None)
self.sampling_rate = hparams.sampling_rate
self.load_mel_from_disk = hparams.load_mel_from_disk
self.return_wavs = opt_get(hparams, ['return_wavs'], False)
@ -70,8 +71,9 @@ class TextMelLoader(torch.utils.data.Dataset):
audio = torch.tensor(audio)
if sampling_rate != self.input_sample_rate:
assert sampling_rate > self.input_sample_rate # Upsampling is not a great idea.
audio = torch.nn.functional.interpolate(audio.unsqueeze(0).unsqueeze(1), scale_factor=self.input_sample_rate/sampling_rate, mode='area')
if sampling_rate < self.input_sample_rate:
print(f'{filename} has a sample rate of {sampling_rate} which is lower than the requested sample rate of {self.input_sample_rate}. This is not a good idea.')
audio = torch.nn.functional.interpolate(audio.unsqueeze(0).unsqueeze(1), scale_factor=self.input_sample_rate/sampling_rate, mode='area')#
audio = (audio.squeeze().clip(-1,1)+1)/2
audio_norm = audio.unsqueeze(0)
audio_norm = torch.autograd.Variable(audio_norm, requires_grad=False)
@ -96,7 +98,12 @@ class TextMelLoader(torch.utils.data.Dataset):
return text_norm
def __getitem__(self, index):
return self.get_mel_text_pair(self.audiopaths_and_text[index])
t, m, p = self.get_mel_text_pair(self.audiopaths_and_text[index])
if self.max_mel_len != None and m.shape[-1] > self.max_mel_len:
# It's hard to handle this situation properly. Best bet is to return the a random valid token and skew the dataset somewhat as a result.
rv = random.randint(0,len(self))
return self[rv]
return t, m, p
def __len__(self):
return len(self.audiopaths_and_text)