From f7d0901ce6e0dbaf1abc44732b3f7b611ebe4cd4 Mon Sep 17 00:00:00 2001
From: James Betker <jbetker@gmail.com>
Date: Sun, 31 Oct 2021 15:01:38 -0600
Subject: [PATCH] Decouple MEL from nv_tacotron_dataset
---
codes/data/__init__.py | 4 +-
codes/data/audio/nv_tacotron_dataset.py | 180 ++++++------------------
2 files changed, 45 insertions(+), 139 deletions(-)
diff --git a/codes/data/__init__.py b/codes/data/__init__.py
index 4feca367..1ede1635 100644
--- a/codes/data/__init__.py
+++ b/codes/data/__init__.py
@@ -61,14 +61,14 @@ def create_dataset(dataset_opt, return_collate=False):
elif mode == 'zipfile':
from data.zip_file_dataset import ZipFileDataset as D
elif mode == 'nv_tacotron':
- from data.audio.nv_tacotron_dataset import TextMelLoader as D
+ from data.audio.nv_tacotron_dataset import TextWavLoader as D
from data.audio.nv_tacotron_dataset import TextMelCollate as C
from models.tacotron2.hparams import create_hparams
default_params = create_hparams()
default_params.update(dataset_opt)
dataset_opt = munchify(default_params)
if opt_get(dataset_opt, ['needs_collate'], True):
- collate = C(dataset_opt.n_frames_per_step)
+ collate = C()
elif mode == 'gpt_tts':
from data.audio.gpt_tts_dataset import GptTtsDataset as D
from data.audio.gpt_tts_dataset import GptTtsCollater as C
diff --git a/codes/data/audio/nv_tacotron_dataset.py b/codes/data/audio/nv_tacotron_dataset.py
index 66e28ae9..a9b22196 100644
--- a/codes/data/audio/nv_tacotron_dataset.py
+++ b/codes/data/audio/nv_tacotron_dataset.py
@@ -6,9 +6,11 @@ import numpy as np
import torch
import torch.utils.data
import torch.nn.functional as F
+import torchaudio
from tqdm import tqdm
import models.tacotron2.layers as layers
+from data.audio.unsupervised_audio_dataset import load_audio
from models.tacotron2.taco_utils import load_wav_to_torch, load_filepaths_and_text
from models.tacotron2.text import text_to_sequence
@@ -37,12 +39,7 @@ def load_voxpopuli(filename):
return filepaths_and_text
-class TextMelLoader(torch.utils.data.Dataset):
- """
- 1) loads audio,text pairs
- 2) normalizes text and converts them to sequences of one-hot vectors
- 3) computes mel-spectrograms from audio files.
- """
+class TextWavLoader(torch.utils.data.Dataset):
def __init__(self, hparams):
self.path = hparams['path']
if not isinstance(self.path, list):
@@ -65,117 +62,67 @@ class TextMelLoader(torch.utils.data.Dataset):
raise NotImplementedError()
self.audiopaths_and_text.extend(fetcher_fn(p))
self.text_cleaners = hparams.text_cleaners
- self.sampling_rate = hparams.sampling_rate
- self.load_mel_from_disk = opt_get(hparams, ['load_mel_from_disk'], False)
- self.return_wavs = opt_get(hparams, ['return_wavs'], False)
- self.input_sample_rate = opt_get(hparams, ['input_sample_rate'], self.sampling_rate)
- assert not (self.load_mel_from_disk and self.return_wavs)
- self.stft = layers.TacotronSTFT(
- hparams.filter_length, hparams.hop_length, hparams.win_length,
- hparams.n_mel_channels, hparams.sampling_rate, hparams.mel_fmin,
- hparams.mel_fmax)
+ self.sample_rate = hparams.sample_rate
random.seed(hparams.seed)
random.shuffle(self.audiopaths_and_text)
- self.max_mel_len = opt_get(hparams, ['max_mel_length'], None)
+ self.max_wav_len = opt_get(hparams, ['max_wav_length'], None)
self.max_text_len = opt_get(hparams, ['max_text_length'], None)
# If needs_collate=False, all outputs will be aligned and padded at maximum length.
self.needs_collate = opt_get(hparams, ['needs_collate'], True)
if not self.needs_collate:
- assert self.max_mel_len is not None and self.max_text_len is not None
+ assert self.max_wav_len is not None and self.max_text_len is not None
- def get_mel_text_pair(self, audiopath_and_text):
+ def get_wav_text_pair(self, audiopath_and_text):
# separate filename and text
audiopath, text = audiopath_and_text[0], audiopath_and_text[1]
text_seq = self.get_text(text)
- mel = self.get_mel(audiopath)
- return (text_seq, mel, text, audiopath_and_text[0])
-
- def get_mel(self, filename):
- if self.load_mel_from_disk and os.path.exists(f'{filename}_mel.npy'):
- melspec = torch.from_numpy(np.load(f'{filename}_mel.npy'))
- assert melspec.size(0) == self.stft.n_mel_channels, (
- 'Mel dimension mismatch: given {}, expected {}'.format(melspec.size(0), self.stft.n_mel_channels))
- else:
- if filename.endswith('.wav'):
- audio, sampling_rate = load_wav_to_torch(filename)
- elif filename.endswith('.mp3'):
- # https://github.com/neonbjb/pyfastmp3decoder - Definitely worth it.
- from pyfastmp3decoder.mp3decoder import load_mp3
- audio, sampling_rate = load_mp3(filename, self.input_sample_rate)
- audio = torch.FloatTensor(audio)
- else:
- audio, sampling_rate = audio2numpy.audio_from_file(filename)
- audio = torch.tensor(audio)
-
- if sampling_rate != self.input_sample_rate:
- 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_norm = torch.nn.functional.interpolate(audio.unsqueeze(0).unsqueeze(1), scale_factor=self.input_sample_rate/sampling_rate, mode='nearest', recompute_scale_factor=False).squeeze()
- else:
- audio_norm = audio
- if audio_norm.std() > 1:
- print(f"Something is very wrong with the given audio. std_dev={audio_norm.std()}. file={filename}")
- return None
- audio_norm.clip_(-1, 1)
- audio_norm = audio_norm.unsqueeze(0)
- audio_norm = torch.autograd.Variable(audio_norm, requires_grad=False)
- if self.input_sample_rate != self.sampling_rate:
- ratio = self.sampling_rate / self.input_sample_rate
- audio_norm = torch.nn.functional.interpolate(audio_norm.unsqueeze(0), scale_factor=ratio, mode='area').squeeze(0)
- if self.return_wavs:
- melspec = audio_norm
- else:
- melspec = self.stft.mel_spectrogram(audio_norm)
- melspec = torch.squeeze(melspec, 0)
-
- return melspec
+ wav = load_audio(audiopath, self.sample_rate)
+ return (text_seq, wav, text, audiopath_and_text[0])
def get_text(self, text):
text_norm = torch.IntTensor(text_to_sequence(text, self.text_cleaners))
return text_norm
def __getitem__(self, index):
- tseq, mel, text, path = self.get_mel_text_pair(self.audiopaths_and_text[index])
- if mel is None or \
- (self.max_mel_len is not None and mel.shape[-1] > self.max_mel_len) or \
+ tseq, wav, text, path = self.get_wav_text_pair(self.audiopaths_and_text[index])
+ if wav is None or \
+ (self.max_wav_len is not None and wav.shape[-1] > self.max_wav_len) or \
(self.max_text_len is not None and tseq.shape[0] > self.max_text_len):
- #if mel is not None:
- # print(f"Exception {index} mel_len:{mel.shape[-1]} text_len:{tseq.shape[0]} fname: {path}")
+ # Basically, this audio file is nonexistent or too long to be supported by the dataset.
# 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.
+ #if wav is not None:
+ # print(f"Exception {index} wav_len:{wav.shape[-1]} text_len:{tseq.shape[0]} fname: {path}")
rv = random.randint(0,len(self)-1)
return self[rv]
- orig_output = mel.shape[-1]
+ orig_output = wav.shape[-1]
orig_text_len = tseq.shape[0]
if not self.needs_collate:
- if mel.shape[-1] != self.max_mel_len:
- mel = F.pad(mel, (0, self.max_mel_len - mel.shape[-1]))
+ if wav.shape[-1] != self.max_wav_len:
+ wav = F.pad(wav, (0, self.max_wav_len - wav.shape[-1]))
if tseq.shape[0] != self.max_text_len:
tseq = F.pad(tseq, (0, self.max_text_len - tseq.shape[0]))
return {
'real_text': text,
'padded_text': tseq,
'input_lengths': torch.tensor(orig_text_len, dtype=torch.long),
- 'padded_mel': mel,
+ 'wav': wav,
'output_lengths': torch.tensor(orig_output, dtype=torch.long),
'filenames': path
}
- return tseq, mel, path, text
+ return tseq, wav, path, text
def __len__(self):
return len(self.audiopaths_and_text)
class TextMelCollate():
- """ Zero-pads model inputs and targets based on number of frames per setep
+ """ Zero-pads model inputs and targets based on number of frames per step
"""
- def __init__(self, n_frames_per_step):
- self.n_frames_per_step = n_frames_per_step
-
def __call__(self, batch):
- """Collate's training batch from normalized text and mel-spectrogram
+ """Collate's training batch from normalized text and wav
PARAMS
------
- batch: [text_normalized, mel_normalized, filename]
+ batch: [text_normalized, wav, filename, text]
"""
# Right zero-pad all one-hot text sequences to max input length
input_lengths, ids_sorted_decreasing = torch.sort(
@@ -193,81 +140,42 @@ class TextMelCollate():
filenames.append(batch[ids_sorted_decreasing[i]][2])
real_text.append(batch[ids_sorted_decreasing[i]][3])
- # Right zero-pad mel-spec
- num_mels = batch[0][1].size(0)
+ # Right zero-pad wav
+ num_wavs = batch[0][1].size(0)
max_target_len = max([x[1].size(1) for x in batch])
- if max_target_len % self.n_frames_per_step != 0:
- max_target_len += self.n_frames_per_step - max_target_len % self.n_frames_per_step
- assert max_target_len % self.n_frames_per_step == 0
# include mel padded and gate padded
- mel_padded = torch.FloatTensor(len(batch), num_mels, max_target_len)
- mel_padded.zero_()
- gate_padded = torch.FloatTensor(len(batch), max_target_len)
- gate_padded.zero_()
+ wav_padded = torch.FloatTensor(len(batch), num_wavs, max_target_len)
+ wav_padded.zero_()
output_lengths = torch.LongTensor(len(batch))
for i in range(len(ids_sorted_decreasing)):
- mel = batch[ids_sorted_decreasing[i]][1]
- mel_padded[i, :, :mel.size(1)] = mel
- gate_padded[i, mel.size(1)-1:] = 1
- output_lengths[i] = mel.size(1)
+ wav = batch[ids_sorted_decreasing[i]][1]
+ wav_padded[i, :, :wav.size(1)] = wav
+ output_lengths[i] = wav.size(1)
return {
'padded_text': text_padded,
'input_lengths': input_lengths,
- 'padded_mel': mel_padded,
- 'padded_gate': gate_padded,
+ 'wav': wav_padded,
'output_lengths': output_lengths,
'filenames': filenames,
'real_text': real_text,
}
-def save_mel_buffer_to_file(mel, path):
- np.save(path, mel.cpu().numpy())
-
-
-def dump_mels_to_disk():
- params = {
- 'mode': 'nv_tacotron',
- 'path': ['Z:\\mozcv\\en\\train.tsv'],
- 'fetcher_mode': ['mozilla_cv'],
- 'phase': 'train',
- 'n_workers': 8,
- 'batch_size': 1,
- 'needs_collate': True,
- 'max_mel_length': 10000,
- 'max_text_length': 1000,
- #'return_wavs': True,
- #'input_sample_rate': 22050,
- #'sampling_rate': 8000
- }
- from data import create_dataset, create_dataloader
- ds, c = create_dataset(params, return_collate=True)
- dl = create_dataloader(ds, params, collate_fn=c)
- for b in tqdm(dl):
- mels = b['padded_mel']
- fnames = b['filenames']
- for j, fname in enumerate(fnames):
- save_mel_buffer_to_file(mels[j], f'{fname}_mel.npy')
-
-
if __name__ == '__main__':
- dump_mels_to_disk()
- '''
+ batch_sz = 32
params = {
'mode': 'nv_tacotron',
- 'path': 'E:\\audio\\MozillaCommonVoice\\en\\train.tsv',
+ 'path': 'E:\\audio\\MozillaCommonVoice\\en\\test.tsv',
'phase': 'train',
- 'n_workers': 12,
- 'batch_size': 32,
+ 'n_workers': 0,
+ 'batch_size': batch_sz,
'fetcher_mode': 'mozilla_cv',
- 'needs_collate': False,
- 'max_mel_length': 800,
- 'max_text_length': 200,
- #'return_wavs': True,
- #'input_sample_rate': 22050,
- #'sampling_rate': 8000
+ 'needs_collate': True,
+ #'max_wav_length': 256000,
+ #'max_text_length': 200,
+ 'sample_rate': 22050,
}
from data import create_dataset, create_dataloader
@@ -277,9 +185,7 @@ if __name__ == '__main__':
m = None
for k in range(1000):
for i, b in tqdm(enumerate(dl)):
- continue
- pm = b['padded_mel']
- pm = torch.nn.functional.pad(pm, (0, 800-pm.shape[-1]))
- m = pm if m is None else torch.cat([m, pm], dim=0)
- print(m.mean(), m.std())
- '''
\ No newline at end of file
+ w = b['wav']
+ for ib in range(batch_sz):
+ print(f'{i} {ib} {b["real_text"][ib]}')
+ torchaudio.save(f'{i}_clip_{ib}.wav', b['wav'][ib], ds.sample_rate)