DL-Art-School/codes/data/audio/nv_tacotron_dataset.py

140 lines
5.0 KiB
Python
Raw Normal View History

import os
import random
import numpy as np
import torch
import torch.utils.data
import models.tacotron2.layers as layers
from models.tacotron2.taco_utils import load_wav_to_torch, load_filepaths_and_text
from models.tacotron2.text import text_to_sequence
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.
"""
def __init__(self, hparams):
self.path = os.path.dirname(hparams['path'])
self.audiopaths_and_text = load_filepaths_and_text(hparams['path'])
self.text_cleaners = hparams.text_cleaners
self.max_wav_value = hparams.max_wav_value
self.sampling_rate = hparams.sampling_rate
self.load_mel_from_disk = hparams.load_mel_from_disk
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)
random.seed(hparams.seed)
random.shuffle(self.audiopaths_and_text)
def get_mel_text_pair(self, audiopath_and_text):
# separate filename and text
audiopath, text = audiopath_and_text[0], audiopath_and_text[1]
audiopath = os.path.join(self.path, audiopath)
text = self.get_text(text)
mel = self.get_mel(audiopath)
return (text, mel)
def get_mel(self, filename):
if not self.load_mel_from_disk:
audio, sampling_rate = load_wav_to_torch(filename)
if sampling_rate != self.stft.sampling_rate:
raise ValueError("{} {} SR doesn't match target {} SR".format(
sampling_rate, self.stft.sampling_rate))
audio_norm = audio / self.max_wav_value
audio_norm = audio_norm.unsqueeze(0)
audio_norm = torch.autograd.Variable(audio_norm, requires_grad=False)
melspec = self.stft.mel_spectrogram(audio_norm)
melspec = torch.squeeze(melspec, 0)
else:
melspec = torch.from_numpy(np.load(filename))
assert melspec.size(0) == self.stft.n_mel_channels, (
'Mel dimension mismatch: given {}, expected {}'.format(
melspec.size(0), self.stft.n_mel_channels))
return melspec
def get_text(self, text):
text_norm = torch.IntTensor(text_to_sequence(text, self.text_cleaners))
return text_norm
def __getitem__(self, index):
return self.get_mel_text_pair(self.audiopaths_and_text[index])
def __len__(self):
return len(self.audiopaths_and_text)
class TextMelCollate():
""" Zero-pads model inputs and targets based on number of frames per setep
"""
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
PARAMS
------
batch: [text_normalized, mel_normalized]
"""
# Right zero-pad all one-hot text sequences to max input length
input_lengths, ids_sorted_decreasing = torch.sort(
torch.LongTensor([len(x[0]) for x in batch]),
dim=0, descending=True)
max_input_len = input_lengths[0]
text_padded = torch.LongTensor(len(batch), max_input_len)
text_padded.zero_()
for i in range(len(ids_sorted_decreasing)):
text = batch[ids_sorted_decreasing[i]][0]
text_padded[i, :text.size(0)] = text
# Right zero-pad mel-spec
num_mels = 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_()
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)
2021-07-09 04:13:44 +00:00
return {
'padded_text': text_padded,
'input_lengths': input_lengths,
'padded_mel': mel_padded,
'padded_gate': gate_padded,
'output_lengths': output_lengths,
2021-07-09 04:13:44 +00:00
}
if __name__ == '__main__':
params = {
'mode': 'nv_tacotron',
'path': 'E:\\audio\\LJSpeech-1.1\\ljs_audio_text_train_filelist.txt',
}
from data import create_dataset
ds = create_dataset(params)
i = 0
m = []
for b in ds:
m.append(b)
i += 1
if i > 9999:
break
m=torch.stack(m)
print(m.mean(), m.std())