-
Notifications
You must be signed in to change notification settings - Fork 9
/
data_loader.py
181 lines (151 loc) · 7.33 KB
/
data_loader.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
from torch.utils import data
from sklearn.preprocessing import StandardScaler
import torch
import glob
import os
from os.path import join, basename, dirname, split, exists
import numpy as np
import h5py
import torchaudio
from utils import to_categorical, load_wav, logmelspectrogram
import json
import librosa
class VQMelSpkEmbDataset(data.Dataset):
def __init__(self, config, split):
super().__init__()
self.min_length = config['min_length']
# load scaler
self.scaler = StandardScaler()
self.scaler.mean_ = np.load(config['stats'])[0]
self.scaler.scale_ = np.load(config['stats'])[1]
self.scaler.n_features_in_ = self.scaler.mean_.shape[0]
# get speakers
self.speakers = json.load(open(config['speakers']))
self.spk_emb_dir = config['spk_emb_dir']
self.spk2files = {}
self.wav_files = []
#parse data_dir
data_dir = config['data_dir']
for spk in self.speakers:
if spk not in self.spk2files:
self.spk2files[spk] = []
if not exists(f'{data_dir}/{spk}'):
raise Exception
_spk_files = sorted(glob.glob(f'{data_dir}/{spk}/E1*.wav')) + sorted(glob.glob(f'{data_dir}/{spk}/E2*.wav'))
if split == 'train':
_spk_files = _spk_files[:-10]
elif split == 'dev':
_spk_files = _spk_files[-10:]
else:
raise Exception
print(spk)
print(len(_spk_files))
self.spk2files[spk].extend(_spk_files)
for f in _spk_files:
self.wav_files.append((spk, f))
self.spk2idx = { spk : ind for ind, spk in enumerate(self.speakers)}
print(f"loading files {len(self.wav_files)}")
self.vq_dir = config['vq_dir']
def __len__(self):
return len(self.wav_files)
def wav_to_mel(self, x):
# load wav
wav = load_wav(x)
mel = logmelspectrogram(wav).T.astype(np.float32)
mel_norm = self.scaler.transform(mel)
return mel_norm
def sample_seg(self, mel, vq = None):
# zero padding
if mel.shape[0] < self.min_length:
mel = np.pad(mel, [[0,self.min_length - mel.shape[0]],[0,0]])
if vq is not None:
vq = np.pad(vq, [[0, self.min_length - vq.shape[0]],[0,0]])
s = np.random.randint(0, mel.shape[0] - self.min_length + 1)
if vq is not None:
return mel[s:s + self.min_length, :], vq[s:s+self.min_length, :]
else:
return mel[s:s+self.min_length, :]
def __getitem__(self, index):
src_spk, src_wav_path = self.wav_files[index]
basename = os.path.basename(src_wav_path).split('.')[0]
src_spk_idx = self.spk2idx[src_spk]
src_spk_cat = np.squeeze(to_categorical([src_spk_idx], num_classes=len(self.speakers)))
src_spk_emb = np.load(os.path.join(self.spk_emb_dir,src_spk+'.npy'))
# sample another source wav
src_ref_wav_idx = np.random.randint(0, len(self.spk2files[src_spk]))
src_ref_wav_path = self.spk2files[src_spk][src_ref_wav_idx]
# sample a target speaker
speakers = self.speakers[:]
# remove source speaker
#speakers.remove(src_spk)
sampled_trg_spk_idx = np.random.randint(0, len(speakers))
trg_spk = speakers[sampled_trg_spk_idx]
trg_spk_emb = np.load(os.path.join(self.spk_emb_dir,trg_spk+'.npy'))
trg_spk_idx = self.spk2idx[trg_spk]
trg_spk_cat = np.squeeze(to_categorical([trg_spk_idx], num_classes=len(self.speakers)))
# sample a target speaker wav
trg_wav_path_idx = np.random.randint(0, len(self.spk2files[trg_spk]))
trg_wav_path = self.spk2files[trg_spk][trg_wav_path_idx]
# extract mels
src_mel = self.wav_to_mel(src_wav_path)
trg_mel = self.wav_to_mel(trg_wav_path)
src_ref_mel = self.wav_to_mel(src_ref_wav_path)
# load vqw2v feat
vqw2v_path=os.path.join(self.vq_dir,src_spk,f'{basename}_dense.npy')
vqw2v_dense = np.load(vqw2v_path).T
if not os.path.exists(vqw2v_path):
raise Exception
# solve mismatch between vq feats and mels
# https://github.com/s3prl/s3prl/blob/master/s3prl/downstream/phone_linear/expert.py
mel_length = src_mel.shape[0]
vq_length = vqw2v_dense.shape[0]
if mel_length > vq_length:
pad_vec = vqw2v_dense[-1,:]
repeated_pad_vec = np.tile(pad_vec, mel_length - vq_length).reshape(mel_length-vq_length,512)
vqw2v_dense = np.concatenate((vqw2v_dense, repeated_pad_vec),0)
elif mel_length < vq_length:
vqw2v_dense = vqw2v_dense[:mel_length,:]
assert src_mel.shape[0] == vqw2v_dense.shape[0], f"mel {src_mel.shape} vq {vqw2v_dense.shape}"
src_mel, vqw2v_dense = self.sample_seg(src_mel, vqw2v_dense)
trg_mel = self.sample_seg(trg_mel)
src_ref_mel = self.sample_seg(src_ref_mel)
# convert to tensor
src_mel_tensor = torch.FloatTensor(src_mel.T).unsqueeze(0)
src_cat_tensor = torch.LongTensor([src_spk_idx]).squeeze_()
src_1hot_tensor = torch.FloatTensor(src_spk_cat)
src_emb_tensor = torch.FloatTensor(src_spk_emb)
trg_mel_tensor = torch.FloatTensor(trg_mel.T).unsqueeze(0)
trg_cat_tensor = torch.LongTensor([trg_spk_idx]).squeeze_()
trg_1hot_tensor = torch.FloatTensor(trg_spk_cat)
trg_emb_tensor = torch.FloatTensor(trg_spk_emb)
src_ref_mel_tensor = torch.FloatTensor(src_ref_mel.T).unsqueeze(0)
vq_tensor = torch.FloatTensor(vqw2v_dense.T).unsqueeze(0)
return src_mel_tensor, src_cat_tensor, src_emb_tensor, trg_mel_tensor, trg_cat_tensor, trg_emb_tensor, src_ref_mel_tensor, vq_tensor
class TestDataset(object):
def __init__(self, speakers, data_dir, split):
self.speakers = json.load(open(speakers))
self.data_dir = data_dir
self.spk2idx = { spk : ind for ind, spk in enumerate(self.speakers)}
self.spk2files = {}
for spk in self.speakers:
if spk not in self.spk2files:
self.spk2files[spk] = []
if exists(f"{data_dir}/{spk}_raw/{spk}_{split}"):
files = glob.glob(f"{data_dir}/{spk}_raw/{spk}_{split}/*.h5")
print(spk)
print(len(files))
self.spk2files[spk].extend(files)
def build_data_loader(config):
train_dataset = eval(config['dataset'])(config, split = 'train')
dev_dataset = eval(config['dataset'])(config, split = 'dev')
train_data_loader = data.DataLoader(dataset=train_dataset,
batch_size=config['batch_size'],
shuffle=config['shuffle'],
num_workers=config['num_workers'],
drop_last=config['drop_last'])
dev_data_loader = data.DataLoader(dataset=dev_dataset,
batch_size=config['batch_size'],
shuffle=False,
num_workers=config['num_workers'],
drop_last=False)
return train_data_loader, dev_data_loader