-
Notifications
You must be signed in to change notification settings - Fork 0
/
features.py
1503 lines (1326 loc) · 62.6 KB
/
features.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
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import argparse
import concurrent.futures
import os
import sys
os.environ['FOR_DISABLE_CONSOLE_CTRL_HANDLER'] = '1'
import subprocess
import audioread
from util.open_unmix.test import OpenUnmixManager
os.environ['FOR_DISABLE_CONSOLE_CTRL_HANDLER'] = '1'
import pathlib
from math import ceil
import librosa
import numpy as np
import pandas as pd
from librosa import frames_to_time, stft, magphase
from librosa.core import istft
from config import SR, RAW_DATA_PATH, FEATURES_DATA_PATH, HOP_LENGTH, N_FFT, N_MELS, FMIN, FMAX, POWER, \
VOICE_DETECTION_MODEL_NAME, N_FFT_HPSS_1, N_HOP_HPSS_1, N_FFT_HPSS_2, N_HOP_HPSS_2, SR_HPSS, \
N_MELS_HPSS, MODELS_DATA_PATH, RNN_INPUT_SIZE_VOICE_ACTIVATION, TOP_DB_WINDOWED_MFCC, \
MIN_INTERVAL_LEN_WINDOWED_MFCC, WINDOW_LEN_WINDOWED_MFCC, WINDOW_HOP_WINDOWED_MFCC, makedirs, AVAIL_MEDIA_TYPES, \
MAGPHASE_WINDOW_SIZE, MAGPHASE_HOP_LENGTH, MAGPHASE_SAMPLE_RATE, MAGPHASE_PATCH_SIZE, \
OUNMIX_SAMPLE_RATE, MFCC_N_COEF, MFCC_FFT_WINDOW, MFCC_HOP_LENGTH, FEATURE_EXTRACTOR_NUM_WORKERS
from util.leglaive.audio import ono_hpss, log_melgram
# MFSC
class FeatureExtractor:
feature_name = 'UnnamedFeature'
dependency_feature_name = ''
def __init__(self, x, y, out_path=FEATURES_DATA_PATH, source_path=RAW_DATA_PATH, feature_path=FEATURES_DATA_PATH,
raw_path=RAW_DATA_PATH):
self.x = x
self.y = y
self.out_path = out_path / self.feature_name
self.feature_path = feature_path
self.raw_path = raw_path
makedirs(self.out_path)
self.source_path = source_path
self.new_labels = []
self.trigger_dependency_warnings_if_needed()
self.trigger_dependency_extraction_if_needed()
print('info: extractor initialized with following data {}'.format(self.x))
try:
self.existing_labels = pd.read_csv(self.get_label_file_path())
print('info: found existing label file in {}'.format(self.get_label_file_path()))
except FileNotFoundError:
self.existing_labels = None
def trigger_dependency_extraction_if_needed(self):
if self.dependency_feature_name:
dependency_extractor = AVAILABLE_FEATURES[self.dependency_feature_name]
try:
df = pd.read_csv(
self.feature_path / dependency_extractor.feature_name / dependency_extractor.get_label_file_name())
return
except Exception as e:
print(str(e))
print("didnt found dependency label file in {}".format(
self.feature_path / dependency_extractor.feature_name / dependency_extractor.get_label_file_name()))
exit(-1)
@classmethod
def from_label_file(cls, label_file_path, out_path=FEATURES_DATA_PATH, source_path=RAW_DATA_PATH):
"""
Initiate extractor from label file (csv) that contains
filename and label columns.
:param label_file_path:
:return:
"""
df = pd.read_csv(label_file_path)
filenames = df['filename']
labels = df['label']
return cls(filenames, labels, out_path=out_path, source_path=source_path)
@classmethod
def magic_init(cls, feature_path=FEATURES_DATA_PATH, raw_path=RAW_DATA_PATH,
raw_label_filename='labels.csv'):
"""
Initiate extractor deducting the paths by the extractor definition.
:param label_file_path:
:return:
"""
from features import AVAILABLE_FEATURES
out_path = feature_path
if cls.dependency_feature_name:
# source path is in feature path
dependency_extractor = AVAILABLE_FEATURES[cls.dependency_feature_name]
source_path = feature_path / dependency_extractor.feature_name
label_file_name = dependency_extractor.get_label_file_name()
else:
# source path is raw data path
source_path = raw_path
label_file_name = raw_label_filename
label_path = source_path / label_file_name
print('info: read metadata from {}'.format(label_path))
print('info: init extractor from {} to {}'.format(source_path, out_path))
df = pd.read_csv(label_path)
filenames = df['filename']
labels = df['label']
print('info: got filenames {}'.format(filenames))
return cls(filenames, labels, out_path=out_path, source_path=source_path, feature_path=feature_path,
raw_path=raw_path)
def remove_feature_files(self, feature_filenames=None):
"""
Remove all files product from this extractor.
:return:
"""
if feature_filenames is None:
feature_filenames = np.asarray(self.new_labels)[:, 0]
[os.remove(self.out_path / filename) for filename in feature_filenames]
os.remove(self.out_path / 'labels.{}.csv'.format(self.feature_name))
def trigger_dependency_warnings_if_needed(self):
"""
Sanity Check for expected input for a particular Feature Extractor,
defined by its dependency_feature_name, compared to the path of the raw_path property.
If dependency_feature_name isn't set, it checks that the input path and filenames are
audifiles.
Also compares the stored filenames in self.x for the dependency_feature_name File Format.
:return: (input_warning, filename_warning) [tuple of booleans] True if corresponding warning
was triggered in this method. A warning is printed when this flag is True.
"""
input_path_warning_flag = True \
if (self.dependency_feature_name and self.source_path != (
FEATURES_DATA_PATH / self.dependency_feature_name)) \
or \
(not self.dependency_feature_name and self.source_path != RAW_DATA_PATH) else False
filename_format_warning_flag = True \
if (self.dependency_feature_name and '.{}.npy'.format(self.dependency_feature_name) not in self.x[0]) \
or \
(not self.dependency_feature_name and self.x[0].split('.')[-1] not in AVAIL_MEDIA_TYPES) else False
if self.dependency_feature_name:
# """
# If this parameter is given, the input is a feature in the Feature folder.
# """
if input_path_warning_flag:
print("""warning:
{} Feature source folder is commonly FEATURES_DATA_PATH/{} config
because it need {} feature as source, receeived {} instead.""".format(self.feature_name,
self.dependency_feature_name,
self.dependency_feature_name,
self.source_path))
if filename_format_warning_flag:
print("""warning:
{} Feature source filenames are commonly formatted like <name>.{}.npy, received {} instead
""".format(self.feature_name, self.dependency_feature_name, self.x[0]))
else:
# """
# If this parameter isn't set, the raw_path should be the RAW_DATA_PATH,
# also the files should end in an accepted format.
# """
if input_path_warning_flag:
print('warning: this FeatureExtractor has a modified self.raw_path ({}), '
'but self.dependency_feature_name wasn\'t set. '.format(self.source_path))
print('If this path doesn\t contain any audio files, '
'this extractor will probably fail.'
'Prefer to set RAW_DATA_PATH for using audio files.') if filename_format_warning_flag else None
if filename_format_warning_flag:
print('warning: No self.dependency_feature_name was set, and '
'the parsed filenames (self.x) has an unsupported media type ({}) '.format(
self.x[0].split('.')[-1]))
return input_path_warning_flag, filename_format_warning_flag
def clean_input_references(self):
"""
Remove elements from x and y that doesnt't have a source (raw) file
:param x: from self.x; a list of filenames (str)
:param y: from self.y; a list of labels (str)
:param raw_path: pathlib.Path object of where source files are located
:return: cleansed x and y
"""
new_x = []
new_y = []
for i, x_i in enumerate(self.x):
if os.path.exists(self.source_path / x_i):
y_i = self.y[i]
new_x.append(x_i)
new_y.append(y_i)
self.x = new_x
self.y = new_y
@staticmethod
def get_mean_voice_activation(voice_activation):
"""
:param voice_activation: Array with shape (n_frames, n_steps)
(note: ~200 predictions are made per frame (aka steps); n_steps is variable depending on frame,
and some predictions can contain NaNs)
:return: Array of shape n_frames. Reduced to just one prediction per frame by mean.
"""
reduction_fun = np.mean
# calculate mean for each frame predictions (~218 per frame)
mean_voice_activation = np.asarray(
[reduction_fun(elem) for elem in voice_activation])
mean_voice_activation = np.nan_to_num(
mean_voice_activation) # remove NaNs product of mean of empty frame
return mean_voice_activation
@staticmethod
def process_element(feature_name, new_labels, out_path, source_path, raw_path, **kwargs):
def __process_element(data):
"""
:param x: filename (str)
:param y: label (str)
:return:
"""
# print('prosessing {}'.format(data))
# x = data[0]
# y = data[1]
# # foo
# product = 'foo'
#
# # this is kind-of standard
# FeatureExtractor.save_feature(product, feature_name, out_path, x, y, new_labels)
raise NotImplementedError()
# stub
return __process_element
@staticmethod
def process_elements(feature_name, new_labels, out_path, source_path, raw_path, fun=None, **kwargs):
def __process_elements(data):
for data_element in data:
fun(data_element)
return __process_elements
def _parallel_transform(self, **kwargs):
"""
Extract features in parallel.
:param kwargs:
:return:
"""
self.clean_input_references()
data = np.asarray([self.x, self.y]).swapaxes(0, 1)
data = self.skip_already_proccessed_in_label_file(data)
# define per-item callable to be processed
process_element = self.process_element(
feature_name=self.feature_name,
new_labels=self.new_labels,
out_path=self.out_path,
source_path=self.source_path,
raw_path=self.raw_path,
features_path=self.feature_path,
existing_labels=self.existing_labels,
**kwargs)
try:
with concurrent.futures.ThreadPoolExecutor(max_workers=FEATURE_EXTRACTOR_NUM_WORKERS) as executor:
iterator = executor.map(process_element, data)
list(iterator)
except KeyboardInterrupt:
print('KeyboardInterrupt catched')
except Exception as e:
print('error: in tranform')
print(e)
self.export_new_labels()
raise e
finally:
print('info: exporting extraction meta-data')
self.export_new_labels()
return np.asarray(self.new_labels)
def _sequential_transform(self, **kwargs):
"""
Extract features sequentially.
:param kwargs:
:return:
"""
self.clean_input_references()
data = np.asarray([self.x, self.y]).swapaxes(0, 1)
data = self.skip_already_proccessed_in_label_file(data)
print('info: starting sequential transform on data {}'.format(data))
print('from {} to {}'.format(self.source_path, self.out_path))
# define per-item callable to be processed
process_element = self.process_element(
feature_name=self.feature_name,
new_labels=self.new_labels,
out_path=self.out_path,
source_path=self.source_path,
raw_path=self.raw_path,
features_path=self.feature_path,
existing_labels=self.existing_labels,
**kwargs)
# define collection callable to process whole data
process_elements = self.process_elements(
feature_name=self.feature_name,
new_labels=self.new_labels,
out_path=self.out_path,
source_path=self.source_path,
raw_path=self.raw_path,
features_path=self.feature_path,
existing_labels=self.existing_labels,
fun=process_element, **kwargs)
try:
process_elements(data)
print('info: finished sequential transform, new labels are {}'.format(self.new_labels))
except KeyboardInterrupt:
print('KeyboardInterrupt catched')
except Exception as e:
print('error: in tranform')
print(e)
self.export_new_labels()
print('exception type {}'.format(type(e)))
raise e
finally:
print('info: exporting extraction meta-data')
self.export_new_labels()
return np.asarray(self.new_labels)
def transform(self, parallel=True, **kwargs):
"""
Transform the data given in Labels to the inteded features.
:param parallel:
:param kwargs:
:return:
"""
if parallel:
return self._parallel_transform(**kwargs)
else:
return self._sequential_transform(**kwargs)
@classmethod
def get_label_file_name(cls):
return 'labels.{}.csv'.format(cls.feature_name)
def get_label_file_path(self):
return self.out_path / self.get_label_file_name()
def export_new_labels(self):
if not self.new_labels:
# if new_labels is empty, then transform did nothing
print('warning: {} did not process any data'.format(self))
return
print('info: exporting processed meta-data to label file in {}'.format(self.get_label_file_path()))
df = pd.DataFrame(np.asarray(self.new_labels))
df.columns = ['filename', 'label']
df.to_csv(self.get_label_file_path(), index=False)
@staticmethod
def get_file_name(x, feature_name, ext='npy'):
"""
Feature File System logic is here
:param ext:
:param x:
:param feature_name:
:return:
"""
# this is kind-of standard
name = '.'.join(x.split('.')[:-1])
filename = '{}.{}.{}'.format(name, feature_name, ext)
return filename
@staticmethod
def save_feature(ndarray, feature_name, out_path, x, y, new_labels, filename=None):
"""
Save any numpy object in Feature File System.
:param ndarray:
:param feature_name:
:param out_path:
:param x:
:param filename:
:return:
"""
# this is kind-of standard
filename = filename or FeatureExtractor.get_file_name(x, feature_name)
np.save(out_path / filename, ndarray)
new_labels.append([filename, y])
print('info: {} transformed and saved!'.format(filename))
return filename
@staticmethod
def save_audio(ndarray, feature_name, out_path, x, y, new_labels, filename=None, sr=SR):
"""
Save any numpy object in Feature File System.
:param ndarray:
:param feature_name:
:param out_path:
:param x:
:param filename:
:return:
"""
# this is kind-of standard
filename = filename or FeatureExtractor.get_file_name(x, feature_name, 'wav')
librosa.output.write_wav(out_path / filename, ndarray, sr=sr, norm=True)
new_labels.append([filename, y])
print('info: {} transformed and saved!'.format(filename))
return filename
@staticmethod
def save_mp3(ndarray, sr, feature_name, out_path, x, y, new_labels, mp3_filename=None):
"""
Save any numpy object in Feature File System.
:param ndarray: 2-axis array with samples and channels [samples, channels]
:param feature_name: unused if mp3 filname is specified
:param out_path:
:param x: can be None if mp3_filename is not None; unused if mp3 filname is specified
:param mp3_filename:
:return:
"""
import soundfile as sf
def _save_mp3(source_path, out_path):
cmd = [
'lame',
'--preset',
'insane',
str(source_path),
str(out_path)
]
errno = subprocess.call(cmd)
if errno:
print('{} encoding failed with code'.format(source_path), end=' ')
print(errno)
print('skipping...')
return errno
os.remove(source_path)
return 0
# this is kind-of standard
if mp3_filename is None:
mp3_filename = FeatureExtractor.get_file_name(x, feature_name, 'mp3')
wav_filename = mp3_filename.replace('mp3', 'wav')
sf.write(str(out_path / wav_filename), ndarray, sr) # write wav file
errno = _save_mp3(out_path / wav_filename,
out_path / mp3_filename) # load wav, encode as mp3 and remove wav file
if errno:
# if any error, then keep wav
filename = wav_filename
else:
# non-error clause, then it was successfully exported to mp3
filename = mp3_filename
if new_labels is not None:
new_labels.append([filename, y])
print('info: {} transformed and saved!'.format(filename))
return filename
def skip_already_proccessed_in_label_file(self, data, enable=True):
"""
Check the existing label file in feature out Path,
remove those filenames from data and
add the existing labels
:param self: FeatureExtractor object.
:param data: list of filename, label
:return:
"""
label_path = self.get_label_file_path()
try:
df = pd.read_csv(label_path)
except FileNotFoundError:
return data
filenames = set(df['filename'])
# todo: finish logic
return data
class MelSpectralCoefficientsFeatureExtractor(FeatureExtractor):
feature_name = 'spec'
@staticmethod
def process_element(feature_name, new_labels, out_path, source_path, **kwargs):
def __process_element(data):
"""
:param x: filename (str)
:param y: label (str)
:return:
"""
print('prosessing {}'.format(data))
x = data[0]
y = data[1]
wav, _ = librosa.load(str(source_path / x), sr=SR)
# Normalize audio signal
wav = librosa.util.normalize(wav)
# Get Mel-Spectrogram
melspec = librosa.feature.melspectrogram(wav, sr=SR, n_fft=N_FFT, hop_length=HOP_LENGTH, n_mels=N_MELS,
fmin=FMIN,
fmax=FMAX, power=POWER)
melspec = librosa.power_to_db(melspec).astype(np.float32)
# this is kind-of standard
FeatureExtractor.save_feature(melspec, feature_name, out_path, x, y, new_labels)
return __process_element
class MelCepstralCoefficientsFeatureExtractor(FeatureExtractor):
feature_name = 'mfcc'
@staticmethod
def process_element(feature_name, new_labels, out_path, source_path, **kwargs):
def __process_element(data):
"""
:param x: filename (str)
:param y: label (str)
:return:
"""
print('prosessing {}'.format(data))
x = data[0]
y = data[1]
file_name = FeatureExtractor.get_file_name(x, feature_name)
try:
# try to load if file already exist
np.load(out_path / file_name, allow_pickle=True)
print('info: {} loaded from .npy !'.format(file_name))
new_labels.append([file_name, y])
except FileNotFoundError or OSError or EOFError:
source_file_path = str(source_path / x)
ext = source_file_path.split('.')[-1]
if 'npy' in ext:
wav = np.load(source_file_path)
elif ext in AVAIL_MEDIA_TYPES:
wav, _ = librosa.load(source_file_path, sr=SR)
else:
raise TypeError('source file ext not recognized ({})'.format(source_file_path))
# Normalize audio signal
wav = librosa.util.normalize(wav)
# Get Mel-Spectrogram
mfcc = librosa.feature.mfcc(wav, sr=SR, n_mfcc=MFCC_N_COEF, n_fft=MFCC_FFT_WINDOW,
hop_length=MFCC_HOP_LENGTH)
# this is kind-of standard
FeatureExtractor.save_feature(mfcc, feature_name, out_path, x, y, new_labels)
return __process_element
def transform(self, parallel=False, **kwargs):
if parallel:
raise Exception('error: {} cannot be ran in paralel'.format(self.feature_name))
return super().transform(parallel, **kwargs)
# MFSC
class WindowedMelSpectralCoefficientsFeatureExtractor(FeatureExtractor):
"""
Do not use.
"""
feature_name = 'windowed_spec'
@staticmethod
def process_element(feature_name, new_labels, out_path, source_path, **kwargs):
def __process_element(data):
"""
:param x: filename (str)
:param y: label (str)
:return:
"""
print('prosessing {}'.format(data))
x = data[0]
y = data[1]
# params
# get song and split
wav, _ = librosa.load(str(source_path / x), sr=SR)
intervals = librosa.effects.split(
# todo: split this extractor in two. One for this split, other for the windows
wav,
top_db=TOP_DB_WINDOWED_MFCC
)
# export intervals as new songs (wav)
for interval_idx, interval in enumerate(intervals):
if interval[1] - interval[0] < MIN_INTERVAL_LEN_WINDOWED_MFCC:
# if length is lesser that 1 second, discard interval
continue
number_of_samples = wav.shape[0]
number_of_windows = ceil(number_of_samples / WINDOW_LEN_WINDOWED_MFCC)
for window_idx in range(number_of_windows):
start_idx = (window_idx * WINDOW_HOP_WINDOWED_MFCC)
end_idx = (start_idx + WINDOW_LEN_WINDOWED_MFCC)
window_wav = wav[start_idx:end_idx]
# Get Mel-Spectrogram
melspec = librosa.feature.melspectrogram(window_wav, sr=SR, n_fft=N_FFT, hop_length=HOP_LENGTH,
n_mels=N_MELS,
fmin=FMIN,
fmax=FMAX, power=POWER)
melspec = librosa.power_to_db(melspec).astype(np.float32)
# this is kind-of standard
filename = FeatureExtractor.get_file_name(x, feature_name,
ext='{}-{}.npy'.format(interval_idx, window_idx))
FeatureExtractor.save_feature(melspec, feature_name, out_path, x, y, new_labels, filename)
return __process_element
# def transform(self, parallel=False, **kwargs):
# if parallel:
# raise Exception('error: {} cannot be ran in paralel'.format(self.feature_name))
# return super().transform(parallel, **kwargs)
# Vocal Separation Pipeline
class MagPhaseFeatureExtractor(FeatureExtractor):
feature_name = 'mag_phase'
@staticmethod
def process_element(feature_name, new_labels, out_path, source_path, **kwargs):
"""
Wrapper for actual function __process_elements(data)
:param feature_name:
:param new_labels:
:param out_path:
:param source_path:
:param fun:
:param model_name:
:param kwargs:
:return:
"""
def __process_element(data):
"""
Compute double stage HPSS for the given audio file
extracted from https://github.com/kyungyunlee/ismir2018-revisiting-svd/blob/master/leglaive_lstm/audio_processor.py
:param x: filename (str)
:param y: label (str)
:return: mel_D2_total : concatenated melspectrogram of percussive, harmonic components of double stage HPSS. Shape=(2 * n_bins, total_frames) ex. (80, 2004)
"""
print('processing {}'.format(data))
x_i = data[0]
y_i = data[1]
file_name = FeatureExtractor.get_file_name(x_i, feature_name)
try:
# try to load if file already exist
np.load(out_path / file_name, allow_pickle=True)
print('info: {} loaded from .npy !'.format(file_name))
new_labels.append([file_name, y_i])
except FileNotFoundError or OSError or EOFError:
# OSError and EOFError are raised if file are inconsistent
audio_src, _ = librosa.load(str(source_path / x_i), sr=MAGPHASE_SAMPLE_RATE)
mix_wav_mag, mix_wav_phase = magphase(
stft(
audio_src,
n_fft=MAGPHASE_WINDOW_SIZE,
hop_length=MAGPHASE_HOP_LENGTH
))
# mix_wav_mag = mix_wav_mag[:, START:END] # 513, SR * Duracion en segundos de x_i
# mix_wav_phase = mix_wav_phase[:, START:END] # ~
# mix_wav_mag = mix_wav_mag[1:].reshape(1, 512, 128, 1) # reshape to match train data
array = np.stack((mix_wav_mag, mix_wav_phase))
# stacks the magnitude and phase,
# final shape should be (2, 513 (n_fft/2 + 1), 128 (patchsize), 1 (dummy channels)
# this is kind-of standard
FeatureExtractor.save_feature(array, feature_name, out_path, x_i, y_i, new_labels)
return __process_element
class SingingVoiceSeparationUnetFeatureExtractor(FeatureExtractor):
feature_name = 'svs_unet'
dependency_feature_name = MagPhaseFeatureExtractor.feature_name
@staticmethod
def process_elements(feature_name, new_labels, out_path, source_path, fun=None,
model_name=VOICE_DETECTION_MODEL_NAME,
**kwargs):
"""
Wrapper for actual function __process_elements(data)
:param feature_name:
:param new_labels:
:param out_path:
:param source_path:
:param fun:
:param model_name:
:param kwargs:
:return:
"""
def __process_elements(data):
"""
:param data: shape (#_songs, 2) the axis 1 corresponds to the filename/label pair
:return:
"""
x = data[:, 0]
y = data[:, 1]
print('loaded metadata in {}'.format(data))
from keras.models import load_model
from keras import backend
if len(backend.tensorflow_backend._get_available_gpus()) > 0:
# set gpu number
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
# load mode
loaded_model = load_model(str(MODELS_DATA_PATH / feature_name / 'latest.h5'.format(model_name)))
print("loaded model")
print(loaded_model.summary())
for idx, x_i in enumerate(x):
# this is kind-of standard
y_i = y[idx]
file_name = FeatureExtractor.get_file_name(x_i, feature_name, ext='wav')
try:
# try to load if file already exist
librosa.load(str(out_path / file_name), sr=MAGPHASE_SAMPLE_RATE)
print('info: {} loaded from .npy !'.format(file_name))
new_labels.append([file_name, y_i])
except FileNotFoundError or OSError or EOFError:
# OSError and EOFError are raised if file are inconsistent
# final_shape: (#_hops, #_mel_filters, #_window)
print('info: loading magphase data for {}'.format(x_i))
magphase = np.load(source_path / x_i) # _data, #_coefs, #_samples)
print('info: formatting data')
try:
# magphase shape (2, 513, #_windows (~4000))
assert len(magphase.shape) == 3
# padding the last dim to fit window_size strictly
number_of_windows = ceil(magphase.shape[2] / MAGPHASE_PATCH_SIZE)
padding = number_of_windows * MAGPHASE_PATCH_SIZE - magphase.shape[2]
if padding > 0:
magphase = np.pad(magphase, ((0, 0), (0, 0), (0, padding)), mode='constant')
# discard first coeficient in both
mag = magphase[0, 1:, :] # shape (512, #_windows)
phase = magphase[1, :, :] # shape (513, #_windows)
x = np.array([mag[:, i * MAGPHASE_PATCH_SIZE:(i + 1) + MAGPHASE_PATCH_SIZE] for i in
range(number_of_windows)])
x = x.reshape(-1, int(MAGPHASE_WINDOW_SIZE / 2), MAGPHASE_PATCH_SIZE, 1)
# stack in a batch of size (512, 128)
print('info: predicting')
y_pred = loaded_model.predict(x, verbose=1) # Shape=(total_frames,)
target_pred_mag = np.vstack((np.zeros((128)), y_pred.reshape(512, 128)))
out_wav = istft(
target_pred_mag * phase
# (mix_wav_mag * target_pred_mag) * mix_wav_phase
, win_length=MAGPHASE_WINDOW_SIZE,
hop_length=MAGPHASE_HOP_LENGTH)
FeatureExtractor.save_audio(out_wav, feature_name, out_path, x_i, y_i, new_labels,
sr=MAGPHASE_SAMPLE_RATE)
except MemoryError as e:
print('error: memory error while proccessing {}. Ignoring...'.format(x_i))
print(e)
return __process_elements
def transform(self, parallel=False, **kwargs):
if parallel:
raise Exception('error: {} cannot be ran in paralel'.format(self.feature_name))
return super().transform(parallel, **kwargs)
class SingingVoiceSeparationOpenUnmixFeatureExtractor(FeatureExtractor):
feature_name = 'svs_openunmix'
@staticmethod
def process_elements(feature_name, new_labels, out_path, source_path, fun=None,
model_name=VOICE_DETECTION_MODEL_NAME, existing_labels=None,
**kwargs):
"""
Wrapper for actual function __process_elements(data)
:param feature_name:
:param new_labels:
:param out_path:
:param source_path:
:param fun:
:param model_name:
:param kwargs:
:return:
"""
def __process_elements(data):
"""
:param data: shape (#_songs, 2) the axis 1 corresponds to the filename/label pair
:return:
"""
x = data[:, 0]
y = data[:, 1]
print('loaded metadata in {}'.format(data))
import torch
no_cuda = True # no cabe en mi gpu :c
use_cuda = not no_cuda and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
sr = OUNMIX_SAMPLE_RATE
model_manager = OpenUnmixManager()
for idx, x_i in enumerate(x):
# for each filename in data
# this is kind-of standard
y_i = y[idx]
ext = 'mp3'
file_name = FeatureExtractor.get_file_name(x_i, feature_name, ext=ext)
try:
# try to load if file already exist
print('info: trying to load {}'.format(out_path / file_name))
if existing_labels is not None and file_name in existing_labels['filename'].values:
new_labels.append([file_name, y_i])
continue
librosa.load(str(out_path / file_name), sr=sr)
new_labels.append([file_name, y_i])
except (FileNotFoundError, OSError, EOFError, audioread.NoBackendError):
# OSError and EOFError are raised if file are inconsistent
# final_shape: (#_hops, #_mel_filters, #_window)
print('info: processing {}'.format(x_i))
audio, rate = librosa.core.load(source_path / x_i, mono=False, sr=sr)
SingingVoiceSeparationOpenUnmixFeatureExtractor.process_x_i(
device,
file_name,
x_i, # as filename is specified manually is just useful for logs
y_i,
source_path,
out_path,
new_labels,
sr,
audio,
rate,
model_manager
)
return __process_elements
@staticmethod
def process_x_i(device, mp3_file_name, x_i, y_i, source_path, out_path, new_labels, sr, audio, rate, model_manager):
"""
:param device:
:param mp3_file_name:
:param x_i: source filename, just useful for logs
:param y_i: label
:param source_path:
:param out_path:
:param new_labels:
:param sr: expected rate
:param audio: audio wav
:param rate: audio sample rate
:return:
"""
# try:
print('info: separating wav with {}'.format(audio.shape))
estimates = model_manager.separate_wav(audio, rate, device)
vocal_wav = estimates['vocals']
FeatureExtractor.save_mp3(
vocal_wav, sr,
feature_name,
out_path,
None,
y_i,
new_labels,
mp3_filename=mp3_file_name
)
# except RuntimeError as e:
# if 'not enough memory' in e.args[0]:
# print(e)
# else:
# raise e
# except MemoryError as e:
# print(e)
# except Exception as e:
# raise e
# finally:
# # this block should be reached only in case of MemoryError or RuntimeError triggered by not enough ram
# print('error: memory error while proccessing {}. Splitting and retrying...'.format(x_i))
# # split & retry logic
# first_filename = mp3_file_name.replace('mp3', '0.mp3')
# second_filename = mp3_file_name.replace('mp3', '1.mp3')
# if audio.shape[1] > 2:
# SingingVoiceSeparationOpenUnmixFeatureExtractor.process_x_i(device, first_filename, None, y_i,
# source_path, out_path, new_labels, sr,
# audio[:, :(audio.shape[1] // 2)], rate,
# model_manager)
# SingingVoiceSeparationOpenUnmixFeatureExtractor.process_x_i(device, second_filename, None, y_i,
# source_path, out_path, new_labels, sr,
# audio[:, (audio.shape[1] // 2):], rate,
# model_manager)
# else:
# print('warning: split and retry reached recursive limit. skipping...')
# return
def transform(self, parallel=False, **kwargs):
if parallel:
raise Exception('error: {} cannot be ran in paralel'.format(self.feature_name))
return super().transform(parallel, **kwargs)
class IntensitySplitterFeatureExtractor(FeatureExtractor):
feature_name = 'intensity_split'
dependency_feature_name = SingingVoiceSeparationOpenUnmixFeatureExtractor.feature_name
@staticmethod
def process_element(feature_name, new_labels, out_path, source_path, **kwargs):
def __process_element(data):
"""
:param x: filename (str)
:param y: label (str)
:return:
"""
print('processing {}'.format(data))
x = data[0]
y = data[1]
filename = FeatureExtractor.get_file_name(x, feature_name,
ext='mp3')
if os.path.isfile(out_path / filename):
new_labels.append([filename, y])
return
# get song and split
wav, _ = librosa.load(str(source_path / x), sr=SR)
intervals = librosa.effects.split(
wav,
top_db=TOP_DB_WINDOWED_MFCC
)
wav_intervals = []
# export intervals as new songs (wav)
for interval_idx, interval in enumerate(intervals):
if interval[1] - interval[0] < MIN_INTERVAL_LEN_WINDOWED_MFCC:
# if length is lesser that 1 second, discard interval
continue
print('debug: appending interval {}'.format(interval))
wav_intervals.append(wav[interval[0]:interval[1]])
FeatureExtractor.save_mp3(
np.concatenate(wav_intervals), SR,
feature_name,
out_path, x, y, new_labels, mp3_filename=filename)
return __process_element
# Singing Voice Detection Pipeline
class DoubleHPSSFeatureExtractor(FeatureExtractor):
feature_name = '2hpss'
dependency_feature_name = SingingVoiceSeparationOpenUnmixFeatureExtractor.feature_name
@staticmethod
def process_element(feature_name, new_labels, out_path, source_path, **kwargs):
"""
Wrapper for actual function __process_elements(data)
:param feature_name:
:param new_labels:
:param out_path:
:param source_path:
:param fun:
:param model_name:
:param kwargs:
:return:
"""
def __process_element(data):
"""
Compute double stage HPSS for the given audio file
extracted from https://github.com/kyungyunlee/ismir2018-revisiting-svd/blob/master/leglaive_lstm/audio_processor.py
:param x: filename (str)
:param y: label (str)
:return: mel_D2_total : concatenated melspectrogram of percussive, harmonic components of double stage HPSS. Shape=(2 * n_bins, total_frames) ex. (80, 2004)
"""
print('processing {}'.format(data))
x_i = data[0]
y_i = data[1]
file_name = FeatureExtractor.get_file_name(x_i, feature_name)
try:
# try to load if file already exist
np.load(out_path / file_name, allow_pickle=True)
print('info: {} loaded from .npy !'.format(file_name))
new_labels.append([file_name, y_i])
except FileNotFoundError or OSError or EOFError:
# OSError and EOFError are raised if file are inconsistent
audio_src, _ = librosa.load(str(source_path / x_i), sr=SR_HPSS)
# Normalize audio signal
audio_src = librosa.util.normalize(audio_src)
# first HPSS
D_harmonic, D_percussive = ono_hpss(audio_src, N_FFT_HPSS_1, N_HOP_HPSS_1)
# second HPSS
D2_harmonic, D2_percussive = ono_hpss(D_percussive, N_FFT_HPSS_2, N_HOP_HPSS_2)
# compute melgram
mel_harmonic = log_melgram(D2_harmonic, SR_HPSS, N_FFT_HPSS_2, N_HOP_HPSS_2, N_MELS_HPSS)
mel_percussive = log_melgram(D2_percussive, SR_HPSS, N_FFT_HPSS_2, N_HOP_HPSS_2, N_MELS_HPSS)
# concat
mel_total = np.vstack((mel_harmonic, mel_percussive))
# this is kind-of standard
FeatureExtractor.save_feature(mel_total, feature_name, out_path, x_i, y_i, new_labels)
return __process_element