-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
488 lines (415 loc) · 15.4 KB
/
train.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
import tensorflow as tf
import argparse
import pprint
import numpy as np
from collections import OrderedDict
from pathlib import Path
from evasion_attack.attack import get_assets, save_npz
from evasion_attack.callbacks import Callbacks
from evasion_attack.centroids import Centroids
from evasion_attack.checkpoints import Checkpoints
from evasion_attack.dataset import IdentificationDataLoader, VerificationDataLoader
from evasion_attack.evaluate import EvaluateIdentificationModel, EvaluateVerificationModel
from evasion_attack.inference import InferenceIdentificationModel, InferenceVerificationModel
from evasion_attack.optimizer import Optimizers
from evasion_attack.utils import set_gpu_growthable, save_config
from evasion_attack.models.preprocess import Preprocessing
from evasion_attack.models.resnet import embedding_model
from evasion_attack.models.trainer import AngularPrototypicalModel
def define_argparser():
""" Define argumemts.
"""
p = argparse.ArgumentParser()
## Default.
p.add_argument(
"--seed",
type=int,
default=42,
help="Arbitrary seed value for reproducibility. Default=%(default)s",
)
p.add_argument(
"--model_type",
type=str,
default="iden", ## or "veri"
choices=["iden", "veri"],
help="The type of model you want to train (iden or veri).Default=%(default)s",
)
p.add_argument(
"--model_name",
type=str,
default=None,
help="The model name. Default=%(default)s",
)
p.add_argument(
"--train_model",
dest="train_model",
action="store_true",
)
p.add_argument(
"--no-train_model",
dest="train_model",
action="store_false",
)
p.set_defaults(
train_model=True
)
p.add_argument(
"--clear_assets",
type=bool,
default=True,
help="Whether to initialize existing checkpoints and log records. Default=%(default)s",
)
## Path.
p.add_argument(
"--data_path",
type=str,
default="data",
help="Folder where initial data is stored. Default=%(default)s",
)
## TFRecord Dataset.
p.add_argument(
"--global_batch_size",
type=int,
default=64,
help="The size of each mini-batch. Default=%(default)s",
)
## Modeling.
p.add_argument(
"--embedding_dim",
type=int,
default=512,
help="Demensions of embedded features. Default=%(default)s",
)
## Training hyper-parameters.
p.add_argument(
"--epochs",
type=int,
default=80, ## 80
help="The number of epochs you want to train with. Default=%(default)s",
)
p.add_argument(
"--init_lr",
type=float,
default=1e-3,
help="Initial learning rate. Warm-up does not proceed. Default=%(default)s",
)
p.add_argument(
"--alpha",
type=float,
default=1./20, ## min_lr = lr * alpha = 5e-5
help="The coefficient that determines the minimum learning rate. min_lr = init_lr * alpha. Default=%(default)s",
)
p.add_argument(
"--rectify",
type=bool,
default=True,
help="Whether to apply rectify in optimizer. Default=%(default)s",
)
p.add_argument(
"--weight_decay",
type=float,
default=5e-4,
help="Size of weight decay. Default=%(default)s",
)
## Callbacks.
p.add_argument(
"--checkpoint_callback",
type=bool,
default=True,
help="Whether to save checkpoints. Default=%(default)s",
)
p.add_argument(
"--tensorboard_callback",
type=bool,
default=True,
help="Whether to log for tensorboard visualizations. Default=%(default)s",
)
p.add_argument(
"--learning_rate_schedular_callback",
type=bool,
default=True,
help="Whether to apply a callback for adjusting the learning rate schedular. Default=%(default)s",
)
p.add_argument(
"--csv_logger_callback",
type=bool,
default=True,
help="Whether to leave log records in csv format. Default=%(default)s",
)
config = p.parse_args()
## Add additional arguments.
config.__setattr__("sample_rate", 16_000)
config.__setattr__("slice_len_sec", 2)
config.__setattr__("slice_len", config.sample_rate * config.slice_len_sec)
config.__setattr__("num_slice", 10)
config.__setattr__("buffer_size", 150_000)
config.__setattr__("tr_folder", str(Path(config.data_path, "vox1_dev_wav", "wav")))
config.__setattr__("ts_folder", str(Path(config.data_path, "vox1_test_wav", "wav")))
config.__setattr__("tfrec_folder", str(Path(config.data_path, "tfrecord")))
config.__setattr__("iden_tfrec_folder", str(Path(config.tfrec_folder, "iden")))
config.__setattr__("veri_tfrec_folder", str(Path(config.tfrec_folder, "veri")))
config.__setattr__("ckpt_dir", "ckpt")
config.__setattr__("log_dir", "logs")
config.__setattr__("result_path", "result")
config.__setattr__("num_classes_for_iden", 1_251)
config.__setattr__("num_classes_for_veri", 1_211)
config.__setattr__("iden_model_name", "AngularPrototypicalModel-Identification")
config.__setattr__("veri_model_name", "AngularPrototypicalModel-Verification")
config.__setattr__("num_iden_ts_ds", 8_251)
config.__setattr__("num_veri_ts_ds", 37_720)
config.__setattr__("attack_type", ["fgm", "pgd"])
config.__setattr__("epsilon", [1e-3, 1e-2, 1e-1])
return config
def build_dataset(config):
""" Build dataset for identification or verification task tricky.
"""
assert config.model_type.lower() in ["iden", "veri"]
def _build_iden_dataset():
""" Build dataset for identification task.
"""
## Load tfrecords and build dataset.
tr_filenames = sorted(list(Path(config.iden_tfrec_folder).glob("tr_*.tfrec")))
vl_filenames = sorted(list(Path(config.iden_tfrec_folder).glob("vl_*.tfrec")))
ts_filenames = sorted(list(Path(config.iden_tfrec_folder).glob("ts_*.tfrec")))
tr_ds = IdentificationDataLoader.get_dataset(
tfrecord_filenames=tr_filenames,
cache=False,
repeats=False,
random_slice=True,
slice_len=config.slice_len_sec * config.sample_rate,
num_slice=config.num_slice,
shuffle=False,
buffer_size=config.buffer_size,
global_batch_size=config.global_batch_size,
)
vl_ds = IdentificationDataLoader.get_dataset(
tfrecord_filenames=vl_filenames,
cache=False,
repeats=False,
random_slice=True,
slice_len=config.slice_len_sec * config.sample_rate,
num_slice=config.num_slice,
shuffle=False,
buffer_size=config.buffer_size,
global_batch_size=config.global_batch_size,
)
ts_ds = IdentificationDataLoader.get_dataset(
tfrecord_filenames=ts_filenames,
cache=False,
repeats=False,
random_slice=False,
slice_len=config.slice_len_sec * config.sample_rate,
num_slice=config.num_slice,
shuffle=False,
buffer_size=config.buffer_size,
global_batch_size=None,
)
return tr_ds, vl_ds, ts_ds
def _build_veri_dataset():
""" Build dataset for verification task.
"""
## Load tfrecords and build dataset.
tr_filenames = sorted(list(Path(config.veri_tfrec_folder).glob("tr_*.tfrec")))
vl_filenames = sorted(list(Path(config.veri_tfrec_folder).glob("vl_*.tfrec")))
ts_filenames = sorted(list(Path(config.veri_tfrec_folder).glob("ts_*.tfrec")))
tr_ds = IdentificationDataLoader.get_dataset( ## not VerificationDataLoader
tfrecord_filenames=tr_filenames,
cache=False,
repeats=False,
random_slice=True,
slice_len=config.slice_len_sec * config.sample_rate,
num_slice=config.num_slice,
shuffle=False,
buffer_size=config.buffer_size,
global_batch_size=config.global_batch_size,
)
vl_ds = IdentificationDataLoader.get_dataset( ## not VerificationDataLoader
tfrecord_filenames=vl_filenames,
cache=False,
repeats=False,
random_slice=True,
slice_len=config.slice_len_sec * config.sample_rate,
num_slice=config.num_slice,
shuffle=False,
buffer_size=config.buffer_size,
global_batch_size=config.global_batch_size,
)
ts_ds = VerificationDataLoader.get_dataset(
tfrecord_filenames=ts_filenames,
cache=False,
repeats=False,
random_slice=False,
slice_len=config.slice_len_sec * config.sample_rate,
num_slice=config.num_slice,
shuffle=False,
buffer_size=config.buffer_size,
global_batch_size=None,
)
return tr_ds, vl_ds, ts_ds
tr_ds, vl_ds, ts_ds = _build_iden_dataset() if config.model_type.lower() == "iden" else _build_veri_dataset()
## Priht the shapes.
print(f"tr_ds: {tr_ds}")
print(f"vl_ds: {vl_ds}")
print(f"ts_ds: {ts_ds}")
return tr_ds, vl_ds, ts_ds
def build_model(config):
sample_rate_ms = int(config.sample_rate / 1_000)
num_classes = config.num_classes_for_iden if config.model_type.lower() == "iden" else config.num_classes_for_veri
## Naming.
if config.model_name == None:
config.model_name = config.iden_model_name if config.model_type == "iden" else config.veri_model_name
## Define the parts.
header = Preprocessing(
frame_length=25 * sample_rate_ms,
frame_step=10 * sample_rate_ms,
fft_length=512,
pad_end=True,
num_mel_bins=64,
sample_rate=config.sample_rate,
lower_edge_hertz=0.,
upper_edge_hertz=8_000.,
)
emb_model = embedding_model(
input_shape=(config.slice_len_sec * config.sample_rate,),
num_classes=num_classes,
embedding_dim=config.embedding_dim,
preprocessing_fn=header,
)
centroids = Centroids(
num_classes=num_classes,
embedding_dim=config.embedding_dim,
)
model = AngularPrototypicalModel(
embedding_model=emb_model,
centroids=centroids,
name=config.model_name,
)
return model
def get_callbacks(config):
""" Get all callbacks.
"""
callbacks = list()
if config.checkpoint_callback:
callbacks.append(Callbacks.get_checkpoint_callback(
ckpt_dir=str(Path(config.ckpt_dir, config.model_type)),
clear_assets=config.clear_assets,
))
if config.tensorboard_callback:
callbacks.append(Callbacks.get_tensorboard_callback(
log_dir=str(Path(config.log_dir, "fit")),
model_name=config.model_name,
clear_assets=config.clear_assets,
))
if config.learning_rate_schedular_callback:
callbacks.append(Callbacks.get_learning_rate_schedular_callback(
init_lr=config.init_lr,
epochs=config.epochs,
alpha=config.epochs,
))
if config.csv_logger_callback:
callbacks.append(Callbacks.get_csv_logger_callback(
log_dir=Path(config.log_dir, "csv"),
model_name=config.model_name,
clear_assets=config.clear_assets,
))
return callbacks
def get_latest_model(config, model: tf.keras.Model):
""" Make clean and get latest model.
"""
Checkpoints.make_clean(
ckpt_dir=str(Path(config.ckpt_dir, config.model_type)),
)
latest_model = Checkpoints.load_latest_checkpoint(
ckpt_dir=str(Path(config.ckpt_dir, config.model_type)),
model=model,
)
return latest_model
def get_prediction(config, latest_model: tf.keras.Model, ts_ds: tf.data.Dataset, save: bool):
""" Get 'y_true' and 'y_pred'.
"""
if config.model_type.lower() == "iden":
total = config.num_iden_ts_ds
y_true, y_pred = InferenceIdentificationModel.inference_fixed_sliced_dataset(
latest_model=latest_model,
ts_ds=ts_ds,
total=total,
)
else:
total = config.num_veri_ts_ds
y_true, y_pred = InferenceVerificationModel.inference_fixed_sliced_dataset(
latest_model=latest_model,
ts_ds=ts_ds,
total=total,
)
## Print the shapes.
print(f"y_true.shape: {y_true.shape}, y_pred.shape: {y_pred.shape}")
## Save the results.
if save:
assets = get_assets(model_type=config.model_type, ds_type="fixed", attack_type=None, epsilon=None)
assets.update({"y_true": y_true, "y_pred": y_pred, "dB_x_delta": None})
save_npz(assets, save_dir=config.result_path)
return y_true, y_pred
def get_evaluation(config, y_true: np.ndarray, y_pred: np.ndarray):
""" Do evaluate performance.
"""
if config.model_type.lower() == "iden":
foo = EvaluateIdentificationModel.cmc(y_true, y_pred)
print(f"Top-1 accuracy: {foo[0]:.4f}, Top-5 accuracy: {foo[4]:.4f}")
else:
foo = EvaluateVerificationModel.eer(y_true, y_pred)
bar = EvaluateVerificationModel.auroc(y_true, y_pred)
qux = EvaluateVerificationModel.min_dcf(y_true, y_pred)
print(f"EER: {foo:.2f}, AUROC: {bar:.4f}, minDCF: {qux:.4f}")
def main(config):
""" Main body.
"""
def print_config(config):
## 'sort_dicts=False' params can only apply python>=3.8.
pprint.PrettyPrinter(indent=4).pprint(OrderedDict(vars(config)))
print_config(config)
## Set gpu memory growthable.
set_gpu_growthable()
## Load tfrecords and build dataset.
tr_ds, vl_ds, ts_ds = build_dataset(config)
## Modeling.
## - total params: 8_082_750
## - trainable params: 7_432_219
## - non-trainable params: 650_531
model = build_model(config)
## After forcibly building the model, print the number of parameters.
model.build([config.global_batch_size, config.slice_len_sec * config.sample_rate])
model.summary()
if config.train_model:
## Compile.
model.compile(
## Custom arguments.
ds=tr_ds,
loss_fn=tf.keras.metrics.Mean(name="loss"),
metric_fn=tf.keras.metrics.SparseCategoricalAccuracy(name="acc"),
## Original arguments.
optimizer=Optimizers.get_adabelief_optimizer(
init_lr=config.init_lr,
rectify=config.rectify,
weight_decay=config.weight_decay,
),
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
)
## Fit.
_ = model.fit(
tr_ds,
validation_data=vl_ds,
epochs=config.epochs,
verbose=1,
callbacks=get_callbacks(config),
)
## Make clean checkpoints and load latest version.
latest_model = get_latest_model(config, model)
## Inference with fixed-sliced dataset and save it.
y_true, y_pred = get_prediction(config, latest_model, ts_ds, save=True)
## Evaluate with some performance.
get_evaluation(config, y_true, y_pred)
## Save configuration.
save_config(vars(config), file_path=Path("config", "train.json"))
if __name__ == "__main__":
config = define_argparser()
main(config)