Spaces:
Runtime error
Runtime error
File size: 7,853 Bytes
3c10b34 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 |
import random
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
import numpy as np
import pandas as pd
import torch
import torch.nn.functional as F
import torchaudio as ta
from einops import pack
from lightning import LightningDataModule
from torch.utils.data.dataloader import DataLoader
from diff_ttsg.text import cmudict, text_to_sequence
from diff_ttsg.text.symbols import symbols
from diff_ttsg.utils.audio import mel_spectrogram
from diff_ttsg.utils.model import fix_len_compatibility, normalize
from diff_ttsg.utils.utils import intersperse, parse_filelist
class CormacDataModule(LightningDataModule):
def __init__(
self,
train_filelist_path,
valid_filelist_path,
batch_size,
num_workers,
pin_memory,
cmudict_path,
motion_folder,
add_blank,
n_fft,
n_feats,
sample_rate,
hop_length,
win_length,
f_min,
f_max,
data_statistics,
motion_pipeline_filename,
seed
):
super().__init__()
# this line allows to access init params with 'self.hparams' attribute
# also ensures init params will be stored in ckpt
self.save_hyperparameters(logger=False)
def setup(self, stage: Optional[str] = None):
"""Load data. Set variables: `self.data_train`, `self.data_val`, `self.data_test`.
This method is called by lightning with both `trainer.fit()` and `trainer.test()`, so be
careful not to execute things like random split twice!
"""
# load and split datasets only if not loaded already
self.trainset = TextMelDataset(
self.hparams.train_filelist_path,
self.hparams.cmudict_path,
self.hparams.motion_folder,
self.hparams.add_blank,
self.hparams.n_fft,
self.hparams.n_feats,
self.hparams.sample_rate,
self.hparams.hop_length,
self.hparams.win_length,
self.hparams.f_min,
self.hparams.f_max,
self.hparams.data_statistics,
self.hparams.seed
)
self.validset = TextMelDataset(
self.hparams.valid_filelist_path,
self.hparams.cmudict_path,
self.hparams.motion_folder,
self.hparams.add_blank,
self.hparams.n_fft,
self.hparams.n_feats,
self.hparams.sample_rate,
self.hparams.hop_length,
self.hparams.win_length,
self.hparams.f_min,
self.hparams.f_max,
self.hparams.data_statistics,
self.hparams.seed
)
def train_dataloader(self):
return DataLoader(
dataset=self.trainset,
batch_size=self.hparams.batch_size,
num_workers=self.hparams.num_workers,
pin_memory=self.hparams.pin_memory,
shuffle=True,
collate_fn=TextMelBatchCollate()
)
def val_dataloader(self):
return DataLoader(
dataset=self.validset,
batch_size=self.hparams.batch_size,
num_workers=self.hparams.num_workers,
pin_memory=self.hparams.pin_memory,
shuffle=False,
collate_fn=TextMelBatchCollate()
)
def teardown(self, stage: Optional[str] = None):
"""Clean up after fit or test."""
pass
def state_dict(self):
"""Extra things to save to checkpoint."""
return {}
def load_state_dict(self, state_dict: Dict[str, Any]):
"""Things to do when loading checkpoint."""
pass
class TextMelDataset(torch.utils.data.Dataset):
def __init__(self, filelist_path, cmudict_path, motion_folder, add_blank=True,
n_fft=1024, n_mels=80, sample_rate=22050,
hop_length=256, win_length=1024, f_min=0., f_max=8000, data_parameters=None, seed=None):
self.filepaths_and_text = parse_filelist(filelist_path)
self.motion_fileloc = Path(motion_folder)
self.cmudict = cmudict.CMUDict(cmudict_path)
self.add_blank = add_blank
self.n_fft = n_fft
self.n_mels = n_mels
self.sample_rate = sample_rate
self.hop_length = hop_length
self.win_length = win_length
self.f_min = f_min
self.f_max = f_max
if data_parameters is not None:
self.data_parameters = data_parameters
else:
self.data_parameters = { 'mel_mean': 0, 'mel_std': 1, 'motion_mean': 0, 'motion_std': 1 }
random.seed(seed)
random.shuffle(self.filepaths_and_text)
def get_pair(self, filepath_and_text):
filepath, text = filepath_and_text[0], filepath_and_text[1]
text = self.get_text(text, add_blank=self.add_blank)
mel = self.get_mel(filepath)
motion = self.get_motion(filepath, mel.shape[1])
return (text, mel, motion)
def get_motion(self, filename, mel_shape, ext=".expmap_86.1328125fps.pkl"):
file_loc = self.motion_fileloc / Path(Path(filename).name).with_suffix(ext)
motion = torch.from_numpy(pd.read_pickle(file_loc).to_numpy())
motion = F.interpolate(motion.T.unsqueeze(0), mel_shape).squeeze(0)
motion = normalize(motion, self.data_parameters['motion_mean'], self.data_parameters['motion_std'])
return motion
def get_mel(self, filepath):
audio, sr = ta.load(filepath)
assert sr == self.sample_rate
mel = mel_spectrogram(audio, self.n_fft, 80, self.sample_rate, self.hop_length,
self.win_length, self.f_min, self.f_max, center=False).squeeze()
mel = normalize(mel, self.data_parameters['mel_mean'], self.data_parameters['mel_std'])
return mel
def get_text(self, text, add_blank=True):
text_norm = text_to_sequence(text, dictionary=self.cmudict)
if self.add_blank:
text_norm = intersperse(text_norm, len(symbols)) # add a blank token, whose id number is len(symbols)
text_norm = torch.IntTensor(text_norm)
return text_norm
def __getitem__(self, index):
text, mel, motion = self.get_pair(self.filepaths_and_text[index])
item = {'y': mel, 'x': text, 'y_motion': motion}
return item
def __len__(self):
return len(self.filepaths_and_text)
def sample_test_batch(self, size):
idx = np.random.choice(range(len(self)), size=size, replace=False)
test_batch = []
for index in idx:
test_batch.append(self.__getitem__(index))
return test_batch
class TextMelBatchCollate(object):
def __call__(self, batch):
B = len(batch)
y_max_length = max([item['y'].shape[-1] for item in batch])
y_max_length = fix_len_compatibility(y_max_length)
x_max_length = max([item['x'].shape[-1] for item in batch])
n_feats = batch[0]['y'].shape[-2]
n_motion = batch[0]['y_motion'].shape[-2]
y = torch.zeros((B, n_feats, y_max_length), dtype=torch.float32)
x = torch.zeros((B, x_max_length), dtype=torch.long)
y_motion = torch.zeros((B, n_motion, y_max_length), dtype=torch.float32)
y_lengths, x_lengths = [], []
for i, item in enumerate(batch):
y_, x_, y_motion_ = item['y'], item['x'], item['y_motion']
y_lengths.append(y_.shape[-1])
x_lengths.append(x_.shape[-1])
y[i, :, :y_.shape[-1]] = y_
x[i, :x_.shape[-1]] = x_
y_motion[i, :, :y_motion_.shape[-1]] = y_motion_
y_lengths = torch.LongTensor(y_lengths)
x_lengths = torch.LongTensor(x_lengths)
return {'x': x, 'x_lengths': x_lengths, 'y': y, 'y_lengths': y_lengths, 'y_motion': y_motion} |