-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset_interface.py
402 lines (283 loc) · 18.1 KB
/
dataset_interface.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
import tensorflow as tf
import numpy as np
features = {
'comb/id': tf.FixedLenFeature([], tf.int64),
'comb/class': tf.FixedLenFeature([], tf.int64),
'comb/genre': tf.FixedLenFeature([], tf.string),
'comb/inst1': tf.FixedLenFeature([], tf.string),
'comb/inst2': tf.FixedLenFeature([], tf.string),
'comb/type1': tf.FixedLenFeature([], tf.string),
'comb/type2': tf.FixedLenFeature([], tf.string),
'comb/file1': tf.FixedLenFeature([], tf.string),
'comb/file2': tf.FixedLenFeature([], tf.string),
'comb/sig1': tf.FixedLenFeature([], tf.string),
'comb/sig2': tf.FixedLenFeature([], tf.string),
'comb/lab1': tf.FixedLenFeature([], tf.string),
'comb/lab2': tf.FixedLenFeature([], tf.string),
'comb/sig1_sample_delay': tf.FixedLenFeature([], tf.int64),
'comb/sig2_sample_delay': tf.FixedLenFeature([], tf.int64),
'comb/ref': tf.FixedLenFeature([], tf.int64),
'comb/label': tf.FixedLenFeature([], tf.int64),
'comb/istrain': tf.FixedLenFeature([], tf.int64),
}
def parse_features_and_decode(tf_example):
parsed_features = tf.parse_single_example(tf_example, features)
parsed_features['comb/sig1'] = tf.reshape(tf.decode_raw(parsed_features['comb/sig1'], tf.float32), [-1])
parsed_features['comb/sig2'] = tf.reshape(tf.decode_raw(parsed_features['comb/sig2'], tf.float32), [-1])
parsed_features['comb/lab1'] = tf.reshape(tf.decode_raw(parsed_features['comb/lab1'], tf.float32), [-1])
parsed_features['comb/lab2'] = tf.reshape(tf.decode_raw(parsed_features['comb/lab2'], tf.float32), [-1])
parsed_features['comb/label'] = tf.cast(parsed_features['comb/label'], tf.int32)
parsed_features['comb/ref'] = tf.cast(parsed_features['comb/ref'], tf.int32)
parsed_features['comb/class'] = tf.cast(parsed_features['comb/class'], tf.int32)
parsed_features['comb/id'] = tf.cast(parsed_features['comb/id'], tf.int32)
parsed_features['comb/sig1_sample_delay'] = tf.cast(parsed_features['comb/sig1_sample_delay'], tf.int32)
parsed_features['comb/sig2_sample_delay'] = tf.cast(parsed_features['comb/sig2_sample_delay'], tf.int32)
parsed_features['comb/istrain'] = tf.cast(parsed_features['comb/istrain'], tf.bool)
return parsed_features
def split_train_test(parsed_features, train_rate):
parsed_features['comb/istrain'] = tf.gather(tf.random_uniform([1], maxval=100, dtype=tf.int32) < tf.cast(train_rate * 100, tf.int32), 0)
return parsed_features
def grab_train_examples(parsed_features):
return parsed_features['comb/istrain']
def grab_test_examples(parsed_features):
return ~parsed_features['comb/istrain']
def turn_into_autotest(parsed_features):
ref = parsed_features['comb/ref']
def delay_positive():
parsed_features['comb/sig2'] = tf.concat([tf.zeros(tf.abs(ref)), parsed_features['comb/sig1'][:-tf.abs(ref)]], axis=0)
parsed_features['comb/lab2'] = tf.concat([tf.zeros(tf.abs(ref)), parsed_features['comb/lab1'][:-tf.abs(ref)]], axis=0)
return parsed_features
def delay_negative():
parsed_features['comb/sig2'] = tf.concat([parsed_features['comb/sig1'][tf.abs(ref):], tf.zeros(tf.abs(ref))], axis=0)
parsed_features['comb/lab2'] = tf.concat([parsed_features['comb/lab1'][tf.abs(ref):], tf.zeros(tf.abs(ref))], axis=0)
return parsed_features
parsed_features = tf.cond(tf.less_equal(ref, 0), delay_positive, delay_negative)
return parsed_features
def add_noiose(parsed_features):
parsed_features['comb/sig1'] = parsed_features['comb/sig1'] + tf.random_normal(tf.shape(parsed_features['comb/sig1']), stddev=1)
parsed_features['comb/sig2'] = parsed_features['comb/sig2'] + tf.random_normal(tf.shape(parsed_features['comb/sig2']), stddev=1)
return parsed_features
def use_activation_signal_instead_of_bit_rate_signal(parsed_features):
parsed_features['comb/sig1'] = parsed_features['comb/lab1']
parsed_features['comb/sig2'] = parsed_features['comb/lab2']
return parsed_features
def filter_split_dataset_from_ids(parsed_features, ids):
id = parsed_features['comb/id']
return tf.reduce_any(tf.equal(id, ids))
def filter_combinations_with_voice(parsed_features):
type1 = parsed_features['comb/type1']
type2 = parsed_features['comb/type2']
return tf.logical_not(tf.logical_or(tf.equal(type1, 'voice'), tf.equal(type2, 'voice')))
def filter_combinations_with_different_types(parsed_features):
type1 = parsed_features['comb/type1']
type2 = parsed_features['comb/type2']
return tf.equal(type1, type2)
def filter_combinations_with_rap_genre(parsed_features):
genre = parsed_features['comb/genre']
return tf.not_equal(genre, 'rap')
def filter_perclass(parsed_features, selected_class):
cls = parsed_features['comb/class']
return tf.reduce_any(tf.equal(cls, selected_class))
def filter_sigsize_leq_N(parsed_features, N):
sig1 = parsed_features['comb/sig1']
sig2 = parsed_features['comb/sig2']
sigsize = tf.minimum(tf.shape(sig1)[0], tf.shape(sig2)[0])
return tf.greater_equal(sigsize, N)
def filter_perwindow(parsed_features, N, nwin, OR):
sig1 = parsed_features['comb/sig1']
sig2 = parsed_features['comb/sig2']
nw1 = 1 + OR * tf.shape(sig1)[0] // N
nw2 = 1 + OR * tf.shape(sig2)[0] // N
return tf.reduce_all([tf.less_equal(nwin + 2, nw1), tf.less_equal(nwin + 2, nw2)])
def clean_from_activation_signal(parsed_features, activation_treshold):
sig1 = parsed_features['comb/sig1']
sig2 = parsed_features['comb/sig2']
lab1 = parsed_features['comb/lab1']
lab2 = parsed_features['comb/lab2']
sigsize = tf.reduce_min([tf.shape(sig1)[0], tf.shape(sig2)[0], tf.shape(lab1)[0], tf.shape(lab2)[0]])
sig1 = sig1[:sigsize]
sig2 = sig2[:sigsize]
lab1 = lab1[:sigsize]
lab2 = lab2[:sigsize]
sigfil1 = tf.reshape(tf.gather(sig1, tf.where(tf.logical_and(lab1 >= activation_treshold, lab2 >= activation_treshold)), axis=0), [-1])
sigfil2 = tf.reshape(tf.gather(sig2, tf.where(tf.logical_and(lab1 >= activation_treshold, lab2 >= activation_treshold)), axis=0), [-1])
parsed_features['comb/sig1'] = sigfil1
parsed_features['comb/sig2'] = sigfil2
return parsed_features
def read_audio_files(audiodir, parsed_features):
wav1 = tf.read_file(tf.string_join([audiodir, parsed_features['comb/file1']], separator='/'))
wav1 = tf.contrib.ffmpeg.decode_audio(wav1, file_format='wav', samples_per_second=44100, channel_count=1)
wav1 = tf.concat([tf.zeros([parsed_features['comb/sig1_sample_delay'], 1]), wav1], axis=0)
wav2 = tf.read_file(tf.string_join([audiodir, parsed_features['comb/file2']], separator='/'))
wav2 = tf.contrib.ffmpeg.decode_audio(wav2, file_format='wav', samples_per_second=44100, channel_count=1)
wav2 = tf.concat([tf.zeros([parsed_features['comb/sig2_sample_delay'], 1]), wav2], axis=0)
parsed_features['sig1/samples'] = tf.divide(wav1, 1.5 * tf.reduce_max(wav1))
parsed_features['sig2/samples'] = tf.divide(wav2, 1.5 * tf.reduce_max(wav2))
return parsed_features
def replace_label_of_unselected_class(parsed_features, selected_class):
label = tf.cast(parsed_features['comb/label'], tf.int32)
cls = tf.cast(parsed_features['comb/class'], tf.int32)
is_from_selected_class = tf.reduce_any(tf.equal(cls, selected_class)) # Check if combination is from class of interest
parsed_features['comb/label'] = tf.cond(is_from_selected_class, lambda: label, lambda: 0)
return parsed_features
def normalize_minmax(sigmat):
max = tf.reduce_max(sigmat, axis=1, keep_dims=True)
min = tf.reduce_min(sigmat, axis=1, keep_dims=True)
signorm = tf.divide(tf.subtract(sigmat, min), tf.subtract(max, min))
return signorm
def normalize_blocks(sigmat):
mean, var = tf.nn.moments(sigmat, axes=[1], keep_dims=True)
signorm = tf.divide(tf.subtract(sigmat, mean), tf.sqrt(var))
return signorm
def gkernel(x, y, s):
return tf.divide(1.0,tf.sqrt(tf.multiply(tf.multiply(2.0,np.pi),s))) * tf.exp( tf.divide(tf.multiply(-1.0,tf.pow(tf.subtract(x,y), 2.0)),tf.multiply(2.0,tf.pow(s, 2.0))) )
def get_most_correlated_blocks(sigmat1, sigmat2, nblocks):
# cy = tf.map_fn(lambda i: gkernel(tf.expand_dims(tf.gather(sigmat1, i, axis=1), dim=1), sigmat2, 0.01), tf.range(tf.shape(sigmat1)[1]), dtype=tf.float32)
cy = tf.map_fn(lambda i: tf.multiply(tf.expand_dims(tf.gather(sigmat1, i, axis=1), dim=1), sigmat2), tf.range(tf.shape(sigmat1)[1]), dtype=tf.float32)
cy = tf.transpose(cy, [1, 0, 2])
cy = tf.reduce_mean(cy, axis=[1, 2])
# cy = tf.reduce_max(cy) - cy
topwin = tf.nn.top_k(cy, nblocks)
return topwin.indices
def prepare_input_with_most_correlated_blocks(parsed_features, N, nwin, OR):
sig1 = parsed_features['comb/sig1']
sig2 = parsed_features['comb/sig2']
sigmat1 = tf.contrib.signal.frame(sig1, N, N // OR, pad_end=False, pad_value=0, axis=-1)
sigmat2 = tf.contrib.signal.frame(sig2, N, N // OR, pad_end=False, pad_value=0, axis=-1)
sigmat1 = normalize_blocks(sigmat1)
sigmat2 = normalize_blocks(sigmat2)
wins = get_most_correlated_blocks(sigmat1, sigmat2, nwin)
sigmat1 = tf.gather(sigmat1, wins, axis=0)
sigmat2 = tf.gather(sigmat2, wins, axis=0)
sigmat1 = tf.where(tf.is_nan(sigmat1), tf.zeros_like(sigmat1), sigmat1)
sigmat2 = tf.where(tf.is_nan(sigmat2), tf.zeros_like(sigmat2), sigmat2)
parsed_features['example/input'] = tf.stack((sigmat1, sigmat2), axis=2)
return parsed_features
def prepare_input_with_sequential_sampling(parsed_features, N, nwin, OR):
sig1 = parsed_features['comb/sig1']
sig2 = parsed_features['comb/sig2']
sigmat1 = tf.contrib.signal.frame(sig1, N, N // OR, pad_end=False, pad_value=0, axis=-1)
sigmat2 = tf.contrib.signal.frame(sig2, N, N // OR, pad_end=False, pad_value=0, axis=-1)
init_id = tf.squeeze(tf.random_uniform([1,], maxval=tf.subtract(tf.shape(sigmat1)[0], nwin), dtype=tf.int32, seed=0))
wins = tf.range(init_id, init_id + nwin)
sigmat1 = tf.gather(sigmat1, wins, axis=0)
sigmat2 = tf.gather(sigmat2, wins, axis=0)
sigmat1 = normalize_blocks(sigmat1)
sigmat2 = normalize_blocks(sigmat2)
sigmat1 = tf.where(tf.is_nan(sigmat1), tf.zeros_like(sigmat1), sigmat1)
sigmat2 = tf.where(tf.is_nan(sigmat2), tf.zeros_like(sigmat2), sigmat2)
parsed_features['example/input'] = tf.stack((sigmat1, sigmat2), axis=2)
return parsed_features
def prepare_input_with_random_sampling(parsed_features, N, nwin, OR):
sig1 = parsed_features['comb/sig1']
sig2 = parsed_features['comb/sig2']
sigmat1 = tf.contrib.signal.frame(sig1, N, N // OR, pad_end=False, pad_value=0, axis=-1)
sigmat2 = tf.contrib.signal.frame(sig2, N, N // OR, pad_end=False, pad_value=0, axis=-1)
# Random window sampling
wins = tf.random_uniform((nwin,), maxval=tf.minimum(tf.shape(sigmat1)[0], tf.shape(sigmat2)[0]), dtype=tf.int32, seed=0)
sigmat1 = tf.gather(sigmat1, wins, axis=0)
sigmat2 = tf.gather(sigmat2, wins, axis=0)
sigmat1 = normalize_blocks(sigmat1)
sigmat2 = normalize_blocks(sigmat2)
sigmat1 = tf.where(tf.is_nan(sigmat1), tf.zeros_like(sigmat1), sigmat1)
sigmat2 = tf.where(tf.is_nan(sigmat2), tf.zeros_like(sigmat2), sigmat2)
parsed_features['example/input'] = tf.stack((sigmat1, sigmat2), axis=2)
return parsed_features
def prepare_input_with_all_windows(parsed_features, N, OR):
sig1 = parsed_features['comb/sig1']
sig2 = parsed_features['comb/sig2']
sigmat1 = tf.contrib.signal.frame(sig1, N, N // OR, pad_end=False, pad_value=0, axis=-1)
sigmat2 = tf.contrib.signal.frame(sig2, N, N // OR, pad_end=False, pad_value=0, axis=-1)
sigmat1 = normalize_blocks(sigmat1)
sigmat2 = normalize_blocks(sigmat2)
parsed_features['example/input'] = tf.stack((sigmat1, sigmat2), axis=2)
return parsed_features
def parse_example(parsed_features):
label = parsed_features['comb/label']
ins = parsed_features['example/input']
inst1 = tf.regex_replace(parsed_features['comb/inst1'], ' ', '_')
inst2 = tf.regex_replace(parsed_features['comb/inst2'], ' ', '_')
type1 = parsed_features['comb/type1']
type2 = parsed_features['comb/type2']
file1 = parsed_features['comb/file1']
file2 = parsed_features['comb/file2']
genre = parsed_features['comb/genre']
id = parsed_features['comb/id']
return ins, label, tf.string_join([type1, ' x ', type2]), tf.string_join([inst1, ' x ', inst2]), genre, id, tf.string_join([file1, ' x ', file2])
def add_defaul_dataset_pipeline(trainParams, modelParams, iterator_handle):
with tf.name_scope('dataset') as scope:
datasetfile = trainParams.dataset_file
# classes = trainParams.selected_class
# np.random.seed(3)
# ncombs = 192401
# eval_ids = np.random.randint(0, ncombs, [np.int32(np.floor(ncombs * 0.35))])
# train_ids = np.setdiff1d(np.array(range(0, ncombs)), eval_ids)
N = modelParams.N
nwin = modelParams.nwin
batch_size = modelParams.batch_size
OR = modelParams.OR
tfdataset = tf.data.TFRecordDataset(datasetfile, compression_type='GZIP', buffer_size=4096)
tfdataset = tfdataset.map(parse_features_and_decode, num_parallel_calls=2)
# tfdataset = tfdataset.map(turn_into_autotest, num_parallel_calls=2) #USE FOR DEBUG ONLY
# tfdataset = tfdataset.map(use_activation_signal_instead_of_bit_rate_signal) #USE FOR DEBUG ONLY
# tfdataset = tfdataset.filter(lambda feat: filter_perclass(feat, classes))
# tfdataset = tfdataset.filter(filter_combinations_with_voice)
# tfdataset = tfdataset.filter(filter_combinations_with_rap_genre)
# tfdataset = tfdataset.filter(filter_combinations_with_different_types)
# tfdataset = tfdataset.map(lambda feat: replace_label_of_unselected_class(feat, classes))
# tfdataset = tfdataset.map(lambda feat: clean_from_activation_signal(feat, 0.5), num_parallel_calls=2)
tfdataset = tfdataset.filter(lambda feat: filter_sigsize_leq_N(feat, N))
# tfdataset = tfdataset.filter(lambda feat: filter_perwindow(feat, N, nwin, OR))
tfdataset = tfdataset.map(lambda feat: prepare_input_with_random_sampling(feat, N, nwin, OR), num_parallel_calls=2)
# tfdataset = tfdataset.map(lambda feat: prepare_input_with_sequential_sampling(feat, N, nwin, OR), num_parallel_calls=2)
# tfdataset = tfdataset.map(lambda feat: prepare_input_with_all_windows(feat, N, OR))
# train_dataset = tfdataset.filter(lambda feat: filter_split_dataset_from_ids(feat, train_ids)).map(parse_example, num_parallel_calls=6)
# test_dataset = tfdataset.filter(lambda feat: filter_split_dataset_from_ids(feat, eval_ids)).map(parse_example, num_parallel_calls=6)
# tfdataset = tfdataset.map(lambda feat: split_train_test(feat, trainParams.train_test_rate), num_parallel_calls=6)
train_dataset = tfdataset.filter(grab_train_examples).map(parse_example, num_parallel_calls=2)
test_dataset = tfdataset.filter(grab_test_examples).map(parse_example, num_parallel_calls=2)
train_dataset = train_dataset.shuffle(4096, reshuffle_each_iteration=True)
test_dataset = test_dataset.shuffle(4096, reshuffle_each_iteration=True)
train_dataset = train_dataset.prefetch(buffer_size=4096)
test_dataset = test_dataset.prefetch(buffer_size=4096)
train_dataset = train_dataset.batch(batch_size)
test_dataset = test_dataset.batch(batch_size)
iterator = tf.data.Iterator.from_string_handle(iterator_handle, train_dataset.output_types, train_dataset.output_shapes)
next_element = iterator.get_next()
train_iterator = train_dataset.make_initializable_iterator()
test_iterator = test_dataset.make_initializable_iterator()
return next_element, train_iterator, test_iterator
def add_augmented_dataset_pipeline(trainParams, modelParams, iterator_handle):
with tf.name_scope('dataset') as scope:
datasetfile = trainParams.dataset_file
N = modelParams.N
nwin = modelParams.nwin
batch_size = modelParams.batch_size
OR = modelParams.OR
tfdataset = tf.data.TFRecordDataset(datasetfile, compression_type='GZIP', buffer_size=4096)
tfdataset = tfdataset.map(parse_features_and_decode, num_parallel_calls=2)
tfdataset = tfdataset.filter(filter_combinations_with_voice)
tfdataset = tfdataset.map(lambda feat: clean_from_activation_signal(feat, 0.5), num_parallel_calls=2)
tfdataset = tfdataset.filter(lambda feat: filter_sigsize_leq_N(feat, N))
real_tr_dataset = tfdataset.filter(grab_train_examples)
auto_tr_dataset = real_tr_dataset.map(turn_into_autotest, num_parallel_calls=4).map(add_noiose, num_parallel_calls=4)
zip_dataset = tf.data.Dataset.zip((auto_tr_dataset, real_tr_dataset))
train_dataset = zip_dataset.flat_map(lambda x0, x1: tf.data.Dataset.from_tensors(x0).concatenate(tf.data.Dataset.from_tensors(x1)))
test_dataset = tfdataset.filter(grab_test_examples)
train_dataset = train_dataset.filter(lambda feat: filter_sigsize_leq_N(feat, N))
test_dataset = test_dataset.filter(lambda feat: filter_sigsize_leq_N(feat, N))
train_dataset = train_dataset.map(lambda feat: prepare_input_with_random_sampling(feat, N, nwin, OR), num_parallel_calls=2)
test_dataset = test_dataset.map(lambda feat: prepare_input_with_random_sampling(feat, N, nwin, OR), num_parallel_calls=2)
train_dataset = train_dataset.map(parse_example, num_parallel_calls=2)
test_dataset = test_dataset.map(parse_example, num_parallel_calls=2)
train_dataset = train_dataset.shuffle(4096, reshuffle_each_iteration=True)
test_dataset = test_dataset.shuffle(4096, reshuffle_each_iteration=True)
train_dataset = train_dataset.prefetch(buffer_size=4096)
test_dataset = test_dataset.prefetch(buffer_size=4096)
train_dataset = train_dataset.batch(batch_size)
test_dataset = test_dataset.batch(batch_size)
iterator = tf.data.Iterator.from_string_handle(iterator_handle, train_dataset.output_types, train_dataset.output_shapes)
next_element = iterator.get_next()
train_iterator = train_dataset.make_initializable_iterator()
test_iterator = test_dataset.make_initializable_iterator()
return next_element, train_iterator, test_iterator