-
Notifications
You must be signed in to change notification settings - Fork 8
/
dataset.py
359 lines (285 loc) · 14.1 KB
/
dataset.py
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
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
from torch.utils import data
import torch
import glob
import os
from os.path import join, basename, dirname, split, exists
import numpy as np
import json
import csv
import random
import librosa
from torch.utils.data import DataLoader
from collections import defaultdict
from prosodic_encoder.ppgvc_f0.ppgvc_lf0 import get_cont_lf0 as process_ppgvc_f0
from prosodic_encoder.fastspeech2_pitch_energy.pitch_energy import process_norm_fastspeech2_pitch_energy
import decoder.vits.commons as vits_commons
def get_dataloader(config):
train_dataset = eval(config['dataset_class'])(config, config['train_meta'], config['train_set'])
dev_dataset = eval(config['dataset_class'])(config, config['dev_meta'], config['dev_set'])
if config['ngpu'] >1:
shuffle = False
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
dev_sampler = torch.utils.data.distributed.DistributedSampler(
dev_dataset)
else:
shuffle = True
train_sampler = None
dev_sampler = None
train_loader = DataLoader(
train_dataset,
batch_size = config['batch_size'],
shuffle = shuffle,
collate_fn = train_dataset.collate_fn,
num_workers = config['num_workers'],
sampler = train_sampler
)
dev_loader = DataLoader(
dev_dataset,
batch_size = config['batch_size'],
shuffle = False,
collate_fn = dev_dataset.collate_fn,
num_workers = config['num_workers'],
sampler = dev_sampler
)
return train_loader, dev_loader
def pad_1D(inputs, length, PAD = 0):
def pad_data(x, length, PAD):
x_padded = np.pad(
x, (0, length - x.shape[0]), mode = 'constant', constant_values = PAD
)
return x_padded
max_len = max(len(x) for x in inputs)
padded = np.stack([pad_data(x, max_len, PAD) for x in inputs])
return padded
def pad_2D(inputs, maxlen = None):
def pad(x, max_len):
PAD = 0
if np.shape(x)[0] > max_len:
raise ValueError(f'shape {x.shape[0]} excceed max_len {max_len}')
s = np.shape(x)[1]
x_padded = np.pad(
x, (0, max_len - x.shape[0]), mode = 'constant', constant_values = PAD
)
return x_padded[:,:s]
if maxlen:
output = np.stack([pad(x,maxlen) for x in inputs])
else:
max_len = max([x.shape[0] for x in inputs])
output = np.stack([pad(x, max_len) for x in inputs])
return output
class VITSDataset(data.Dataset):
def __init__(self, config, metadata_csv, split):
super().__init__()
self.metadata = []
# setup
self.hop_size = config['vits_hop_size']
self.sampling_rate = config['sampling_rate']
self.segment_size = config['decoder_params']['segment_size'] # random slice segment size of the HIfIGAN in the VITS model.
# read metadata
with open(metadata_csv) as f:
reader = csv.DictReader(f, delimiter = ',')
for row in reader:
# remove utterances that are too long for training.
if config['rm_long_utt']:
_duration = row['duration']
if float(_duration) < config['max_utt_duration']:
self.metadata.append(row)
f.close()
print(f'{split} data samples {len(self.metadata)}')
self.batch_size = config['batch_size']
self.drop_last = config['drop_last']
self.sort = config['sort']
# feature dirs
self.spec_dir = os.path.join(config['dump_dir'], config['dataset'], split, config['mel_type'])
self.ling_enc = config['ling_enc']
self.ling_rep_dir = os.path.join(config['dump_dir'], config['dataset'], split, self.ling_enc)
self.spk_enc = config['spk_enc']
self.spk_emb_dir = os.path.join(config['dump_dir'], config['dataset'], split, self.spk_enc)
self.pros_enc = config['pros_enc'] #e.g. ppgvc_f0
self.pros_rep_dir = os.path.join(config['dump_dir'], config['dataset'], split, self.pros_enc)
self.pros_rep_process_func = f'process_{self.pros_enc}'
def __len__(self):
return len(self.metadata)
def __getitem__(self, idx):
row = self.metadata[idx]
ID = row['ID']
spk = row['spk']
wav_path = row['wav_path']
start, end = float(row['start']), float(row['end'])
# audio
audio, fs = librosa.load(wav_path, sr = self.sampling_rate)
audio = audio[ int(start * self.sampling_rate):
int(end * self.sampling_rate)
]
# feature path
spec_path = os.path.join(self.spec_dir, spk, ID + '.npy')
ling_rep_path = os.path.join(self.ling_rep_dir, spk, ID+'.npy')
spk_emb_path = os.path.join(self.spk_emb_dir, spk, ID+'.npy')
pros_rep_path = os.path.join(self.pros_rep_dir, spk, ID + '.npy')
assert os.path.exists(spec_path), f"{spec_path}"
assert os.path.exists(ling_rep_path), f'{ling_rep_path}'
assert os.path.exists(spk_emb_path), f'{spk_emb_path}'
assert os.path.exists(pros_rep_path), f'{pros_rep_path}'
# load feature
spec = np.load(spec_path)
spec_duration = spec.shape[0]
# pad spec to match the segment_size
spec_segment_size = self.segment_size // self.hop_size
if spec_duration < spec_segment_size:
spec_pad_length = spec_segment_size - spec_duration
spec = np.pad(spec, [[0,spec_pad_length],[0,0]], mode = 'constant', constant_values = 0.)
assert spec.shape[0] >= spec_segment_size
spec_duration = spec.shape[0]
audio_duration = audio.shape[0]
ling_rep = np.load(ling_rep_path)
ling_duration = ling_rep.shape[0]
spk_emb = np.load(spk_emb_path)
pros_rep = np.load(pros_rep_path)
pros_rep = eval(self.pros_rep_process_func)(pros_rep)
pros_duration = pros_rep.shape[0]
# up_sample ling_rep to 10hz, in case some ling_rep are 50hz or 25hz.
factor = int(round(spec_duration / ling_duration))
repeated_ling_rep = np.repeat(ling_rep, factor, axis=1)
ling_rep = np.reshape(repeated_ling_rep, [ling_duration * factor, ling_rep.shape[1]])
ling_duration = ling_rep.shape[0]
# match length between spec and ling_rep
if spec_duration > ling_duration :
pad_vec = np.expand_dims(ling_rep[-1,:], axis = 0)
ling_rep = np.concatenate((ling_rep, np.repeat(pad_vec, spec_duration - ling_duration, 0)),0)
elif spec_duration < ling_duration:
ling_rep = ling_rep[:spec_duration,:]
# match length between spec and pros_rep
if spec_duration > pros_duration:
pad_vec = np.expand_dims(pros_rep[-1,:],axis = 0)
pros_rep = np.concatenate((pros_rep, np.repeat(pad_vec, spec_duration - pros_duration, 0)),0)
elif spec_duration < pros_duration:
pros_rep = pros_rep[:spec_duration,:]
# match length between audio and spec
if audio_duration < int(spec_duration * self.hop_size):
#pad
pad_length = int(spec_duration * self.hop_size) - audio_duration
audio = np.concatenate([audio, np.array([0.]*pad_length)], axis = 0)
elif audio_duration > int(spec_duration * self.hop_size):
audio = audio[:int(spec_duration * self.hop_size)]
return (audio, spec, ling_rep, pros_rep, spk_emb, spec_duration)
def collate_fn(self, data):
# sort in batch
batch_size = len(data)
if self.sort:
len_arr = np.array([d[-1] for d in data])
idx_arr = np.argsort(~len_arr)
else:
idx_arr = np.arange(batch_size)
audio = [ data[id][0] for id in idx_arr]
spec = [data[id][1] for id in idx_arr]
ling_rep = [ data[id][2] for id in idx_arr]
pros_rep = [ data[id][3] for id in idx_arr]
spk_emb = [ data[id][4] for id in idx_arr]
spec_length = [ data[id][5] for id in idx_arr ]
audio_length = [len(_audio) for _audio in audio]
max_spec_len = max(spec_length)
max_wav_len = max(audio_length)
padded_audio = torch.FloatTensor(pad_1D(audio, max_wav_len)).unsqueeze(1)
padded_spec = torch.FloatTensor(pad_2D(spec, max_spec_len)).transpose(1,2)
padded_ling_rep = torch.FloatTensor(pad_2D(ling_rep, max_spec_len)).transpose(1,2)
padded_pros_rep = torch.FloatTensor(pad_2D(pros_rep, max_spec_len)).transpose(1,2)
spk_emb_tensor = torch.FloatTensor(np.array(spk_emb)).unsqueeze(2)
spec_length = torch.LongTensor(np.array(spec_length))
audio_length = torch.LongTensor(np.array(audio_length))
output = (padded_audio, padded_spec, padded_ling_rep, padded_pros_rep, spk_emb_tensor, spec_length, audio_length)
return output
class Dataset(data.Dataset):
def __init__(self, config, metadata_csv, split):
super().__init__()
self.metadata = []
# read metadata
with open(metadata_csv) as f:
reader = csv.DictReader(f, delimiter = ',')
for row in reader:
# remove utterances that are too long for training.
if config['rm_long_utt']:
_duration = row['duration']
if float(_duration) < config['max_utt_duration']:
self.metadata.append(row)
f.close()
print(f'{split} data samples {len(self.metadata)}')
self.batch_size = config['batch_size']
self.drop_last = config['drop_last']
self.sort = config['sort']
# feature dirs
self.mel_dir = os.path.join(config['dump_dir'], config['dataset'], split, config['mel_type'])
self.ling_enc = config['ling_enc']
self.ling_rep_dir = os.path.join(config['dump_dir'], config['dataset'], split, self.ling_enc)
self.spk_enc = config['spk_enc']
self.spk_emb_dir = os.path.join(config['dump_dir'], config['dataset'], split, self.spk_enc)
self.pros_enc = config['pros_enc'] #e.g. ppgvc_f0
self.pros_rep_dir = os.path.join(config['dump_dir'], config['dataset'], split, self.pros_enc)
self.pros_rep_process_func = f'process_{self.pros_enc}'
# frames per step (only work for TacoMOL)
self.frames_per_step = config['frames_per_step'] if 'frames_per_step' in config else 1
def __len__(self):
return len(self.metadata)
def __getitem__(self, idx):
row = self.metadata[idx]
ID = row['ID']
spk = row['spk']
# feature path
mel_path = os.path.join(self.mel_dir, spk, ID + '.npy')
ling_rep_path = os.path.join(self.ling_rep_dir, spk, ID+'.npy')
spk_emb_path = os.path.join(self.spk_emb_dir, spk, ID+'.npy')
pros_rep_path = os.path.join(self.pros_rep_dir, spk, ID + '.npy')
assert os.path.exists(mel_path), f"{mel_path}"
assert os.path.exists(ling_rep_path), f'{ling_rep_path}'
assert os.path.exists(spk_emb_path), f'{spk_emb_path}'
assert os.path.exists(pros_rep_path), f'{pros_rep_path}'
# load feature
mel = np.load(mel_path)
mel_duration = mel.shape[0]
ling_rep = np.load(ling_rep_path)
ling_duration = ling_rep.shape[0]
spk_emb = np.load(spk_emb_path)
pros_rep = np.load(pros_rep_path)
pros_rep = eval(self.pros_rep_process_func)(pros_rep)
pros_duration = pros_rep.shape[0]
# up_sample ling_rep to 10hz, in case some ling_rep are 50hz or 25hz.
factor = int(round(mel_duration / ling_duration))
repeated_ling_rep = np.repeat(ling_rep, factor, axis=1)
ling_rep = np.reshape(repeated_ling_rep, [ling_duration * factor, ling_rep.shape[1]])
ling_duration = ling_rep.shape[0]
# match length between mel and ling_rep
if mel_duration > ling_duration :
pad_vec = np.expand_dims(ling_rep[-1,:], axis = 0)
ling_rep = np.concatenate((ling_rep, np.repeat(pad_vec, mel_duration - ling_duration, 0)),0)
elif mel_duration < ling_duration:
ling_rep = ling_rep[:mel_duration,:]
# match length between mel and pros_rep
if mel_duration > pros_duration:
pad_vec = np.expand_dims(pros_rep[-1,:],axis = 0)
pros_rep = np.concatenate((pros_rep, np.repeat(pad_vec, mel_duration - pros_duration, 0)),0)
elif mel_duration < pros_duration:
pros_rep = pros_rep[:mel_duration,:]
return (mel, ling_rep, pros_rep, spk_emb, mel_duration)
def collate_fn(self, data):
# sort in batch
batch_size = len(data)
if self.sort:
len_arr = np.array([d[-1] for d in data])
idx_arr = np.argsort(~len_arr)
else:
idx_arr = np.arange(batch_size)
mel = [ data[id][0] for id in idx_arr]
ling_rep = [ data[id][1] for id in idx_arr]
pros_rep = [ data[id][2] for id in idx_arr]
spk_emb = [ data[id][3] for id in idx_arr]
length = [ data[id][4] for id in idx_arr ]
max_len = max(length)
if max_len % self.frames_per_step != 0:
max_len += (self.frames_per_step - max_len % self.frames_per_step)
padded_mel = torch.FloatTensor(pad_2D(mel, max_len))
padded_ling_rep = torch.FloatTensor(pad_2D(ling_rep, max_len))
padded_pros_rep = torch.FloatTensor(pad_2D(pros_rep, max_len))
spk_emb_tensor = torch.FloatTensor(np.array(spk_emb)).unsqueeze(1)
length = torch.LongTensor(np.array(length))
output = (padded_mel, padded_ling_rep, padded_pros_rep, spk_emb_tensor, length, max_len)
return output