-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdataloader.py
299 lines (255 loc) · 11 KB
/
dataloader.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
import os
import numpy as np
import torch
import torch.utils.data
import librosa
from utils.utils import seed_everything
from utils.prep_utils import stft_process,loudness_normalize
from utils.aug_utils import mix_db
seed_everything(42)
def load_stft_noise(is_test=False):
normal_dir = '../eggdata/TrainSTFT/noise/normal/'
musical_dir = '../eggdata/TrainSTFT/noise/musical/'
normal_noise_files = os.listdir(normal_dir)
musical_noise_files = os.listdir(musical_dir)
if is_test:
normal_noise_files = normal_noise_files[:10]
musical_noise_files = musical_noise_files[:10]
normal_noise = [np.load(normal_dir+file) for file in normal_noise_files]
musical_noise = [np.load(musical_dir+file) for file in musical_noise_files]
return normal_noise, musical_noise
class Dataset(torch.utils.data.Dataset):
def __init__(self,X,n_sample=1024,stride = 512, config = None,is_train=True,aug = None):
self.X = X
self.n_sample = n_sample
self.stride = stride
self.is_train = is_train
self.aug = aug
self.config = config
def __len__(self):
return (self.X.shape[1]-self.n_sample)//self.stride -1
def __getitem__(self,idx):
start_frame = idx*self.stride + np.random.randint(0,self.n_sample-1)
# start_frame = idx*self.stride + self.n_sample-1
X = np.array(self.X[0,start_frame : start_frame + self.n_sample].tolist())
y = np.array(self.X[1,start_frame : start_frame + self.n_sample].tolist())
if self.is_train:
## add white noise
db = np.random.uniform(0,20)
noise = np.random.normal(size = X.shape)
X = mix_db(X,noise,db)
X = loudness_normalize(X)
X = librosa.stft(X,n_fft = self.config["window_length"], hop_length = self.config["hop_length"], window = self.config["window"])
y = librosa.stft(y,n_fft = self.config["window_length"], hop_length = self.config["hop_length"], window = self.config["window"])
X = stft_process(X).astype(np.float32)
y = stft_process(y,mask=True).astype(np.float32)
return X,y
# class Dataset(torch.utils.data.Dataset):
# def __init__(self,X,n_frame,stride,is_train = True,normal_noise = None,musical_noise=None,aug=None,pseudo_mode=False):
# self.data = np.load(X,mmap_mode='r') if isinstance(X,str) else X
# self.n_frame = n_frame
# self.stride = stride
# self.is_train = is_train
# self.normal_noise = normal_noise
# self.musical_noise = musical_noise
# self.aug= aug
# self.pseudo_mode=pseudo_mode
# def __len__(self):
# return (self.data.shape[-1] - self.n_frame - self.stride) // self.stride
# def __getitem__(self,idx):
# if self.stride <=self.n_frame:
# offset = np.random.randint(low=0,high=self.stride)
# [X,Y] = self.data[:,:,offset + idx*self.stride:offset + idx*self.stride + self.n_frame] ## 2,F,T
# else:
# random_idx = np.random.randint(low = idx*self.stride, high=(idx+1)*self.stride - self.n_frame - 1)
# [X,Y] = self.data[:,:,random_idx:random_idx+self.n_frame]
# # Mixing before mel
# if self.aug:
# X = self.aug(X,self.normal_noise,self.musical_noise)
# X = stft_process(X).astype(np.float32) ## 2,F,T mag,phase
# Y = stft_process(Y,mask=True).astype(np.float32) ## 3,F,T mag,phase,mask
# # SpecAug after mel
# if self.is_train:
# X = add_whitenoise(X)
# if self.pseudo_mode:
# return X,Y,True
# return X,Y
## dataset cat 하기
#####
"""
def load_stft_unlabel_datas_path(is_test = False):
TrainPath = '../eggdata/TrainSTFT/unlabeled/'
# TrainPath = '../eggdata/TrainSTFT/unlabeled/split/'
Train = []
dsets = ['DSD100', 'KSS', 'zeroth_korean', 'speech_ko'] ## except librispeech
for dataset in os.listdir(TrainPath):
if dataset in dsets:
for file in os.listdir(TrainPath + dataset):
if 'npy' in file:
Train.append(os.path.join(TrainPath + dataset,file))
if is_test:
Train = Train[:10000]
return Train
def load_stft_datas_path(is_test = False,pseudo = None):
TrainPath = '../eggdata/TrainSTFT/train_data/'
ValidPath = '../eggdata/TrainSTFT/valid_data/'
Train = []
Val = []
for file in os.listdir(TrainPath):
if 'npy' in file:
Train.append(TrainPath + file)
for file in os.listdir(ValidPath):
if 'npy' in file:
Val.append(ValidPath + file)
realNum = len(Train)
if pseudo is not None:
pseudoTrainPath = '../eggdata/TrainSTFT/pseudo_label/%s/'%pseudo
for dataset in os.listdir(pseudoTrainPath):
for file in os.listdir(pseudoTrainPath + dataset):
if 'npy' in file:
Train.append(os.path.join(pseudoTrainPath + dataset,file))
if is_test:
Train = Train[:1000] + Train[-1000:]
Val = Val[:1000] + Val[-1000:]
return Train,Val
def pseudoPath2Path(pseudoPath):
path = '../eggdata/TrainSTFT/unlabeled/' + '/'.join(pseudoPath.split('/')[-2:])
return path
class Dataset(torch.utils.data.Dataset):
def __init__(self,X,n_frame,is_train = True,normal_noise = None,musical_noise=None,aug=None):
self.data = X
self.n_frame = n_frame
self.is_train = is_train
self.normal_noise = normal_noise
self.musical_noise = musical_noise
self.aug= aug
def __len__(self):
return len(self.data)
def __getitem__(self,idx):
if isinstance(self.data[idx],str):
[X,Y] = np.load(self.data[idx])
else:
[X,Y] = self.data[idx]
### random crop
T = X.shape[1]
if T<self.n_frame:
left = np.random.randint(self.n_frame-T)
right = self.n_frame-T-left
X = np.pad(X,((0,0),(left,right)),mode = 'constant',constant_values=(-1))
Y = np.pad(Y,((0,0),(left,right)),mode = 'constant',constant_values=(-1))
elif T>self.n_frame:
pi = np.random.randint(T-self.n_frame)
X = X[:,pi:pi+self.n_frame]
Y = Y[:,pi:pi+self.n_frame]
# Mixing before mel
if self.aug:
X = self.aug(X,self.normal_noise,self.musical_noise)
X = stft_process(X) ## 2,F,T mag,phase
Y = stft_process(Y,mask=True) ## 3,F,T mag,phase,mask
# SpecAug after mel
if self.is_train:
X = add_whitenoise(X)
X = spec_masking(X, F = 5, T = 1, num_masks = 10, prob = 1, replace_with_zero = True)
return X,Y
class STDataset(torch.utils.data.Dataset):
def __init__(self,X,n_frame,is_train = True,normal_noise = None,musical_noise=None,aug=None):
self.data = X
self.n_frame = n_frame
self.is_train = is_train
self.normal_noise = normal_noise
self.musical_noise = musical_noise
self.aug= aug
def __len__(self):
return len(self.data)
def __getitem__(self,idx):
pseudo = 1 if 'pseudo' in self.data[idx] else 0
# a = time.time()
if isinstance(self.data[idx],str):
if pseudo == 0:
[X,Y] = np.load(self.data[idx])
else:
X = np.load(pseudoPath2Path(self.data[idx]))
Y = np.load(self.data[idx])
else:
[X,Y] = self.data[idx]
# print('pseudo :',pseudo==1, time.time()-a)
### random crop
T = X.shape[1]
if T<self.n_frame:
left = np.random.randint(self.n_frame-T)
right = self.n_frame-T-left
X = np.pad(X,((0,0),(left,right)),mode = 'constant',constant_values=(-1))
Y = np.pad(Y,((0,0),(left,right)),mode = 'constant',constant_values=(-1))
elif T>self.n_frame:
pi = np.random.randint(T-self.n_frame)
X = X[:,pi:pi+self.n_frame]
Y = Y[:,pi:pi+self.n_frame] if pseudo == 0 else Y[:,:,pi:pi+self.n_frame]
# Mixing before mel
if self.aug:
X = self.aug(X,self.normal_noise,self.musical_noise)
X = stft_process(X) ## 2,F,T mag,phase
if pseudo == 0:
Y = stft_process(Y,mask=True) ## 3,F,T mag,phase,mask
else:
Y[2,:,:] = np.round(expit(Y[2,:,:]))
# SpecAug after mel
if self.is_train:
X = add_whitenoise(X)
X = spec_masking(X, F = 5, T = 1, num_masks = 10, prob = 1, replace_with_zero = True)
return X,Y,pseudo
class SSLDataset(torch.utils.data.Dataset):
def __init__(self,X,unlabel,n_frame,is_train = True,normal_noise = None,musical_noise=None,aug=None):
self.data = X
self.unlabel = unlabel
self.n_frame = n_frame
self.is_train = is_train
self.normal_noise = normal_noise
self.musical_noise = musical_noise
self.aug= aug
self.data_len = len(self.data)
def __len__(self):
return len(self.data) + len(self.unlabel)
def __getitem__(self,idx):
labeled = 1
if idx < self.data_len:
if isinstance(self.data[idx],str):
[X,Y] = np.load(self.data[idx],allow_pickle=True)
else:
[X,Y] = self.data[idx]
else:
if isinstance(self.unlabel[self.data_len - idx],str):
X = np.load(self.unlabel[self.data_len - idx],allow_pickle=True)
else:
X = self.unlabel[self.data_len - idx]
labeled = 0
### random crop
T = X.shape[1]
if T<self.n_frame:
left = np.random.randint(self.n_frame-T)
right = self.n_frame-T-left
X = np.pad(X,((0,0),(left,right)),mode = 'constant',constant_values=(-1))
if labeled == 1:
Y = np.pad(Y,((0,0),(left,right)),mode = 'constant',constant_values=(-1))
elif T>self.n_frame:
pi = np.random.randint(T-self.n_frame)
X = X[:,pi:pi+self.n_frame]
if labeled == 1:
Y = Y[:,pi:pi+self.n_frame]
# Mixing before mel
if self.aug:
X1 = self.aug(X,self.normal_noise,self.musical_noise)
X2 = self.aug(X,self.normal_noise,self.musical_noise)
X1 = stft_process(X1) ## 2,F,T mag,phase
X2 = stft_process(X2) ## 2,F,T mag,phase
if labeled == 1:
Y = stft_process(Y,mask=True) ## 3,F,T mag,phase,mask
else:
Y = np.zeros((3,257,self.n_frame)).astype(X1.dtype)
# SpecAug after mel
if self.is_train:
X1 = add_whitenoise(X1)
X1 = spec_masking(X1, F = 10, T = 2, num_masks = 3, prob = 1, replace_with_zero = True)
X2 = add_whitenoise(X2)
X2 = spec_masking(X2, F = 10, T = 2, num_masks = 3, prob = 1, replace_with_zero = True)
return (X1,X2),Y, labeled
"""