xvector.py 91.3 KB
Newer Older
Anthony Larcher's avatar
Anthony Larcher committed
1
# coding: utf-8 -*-
Anthony Larcher's avatar
Anthony Larcher committed
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
#
# This file is part of SIDEKIT.
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is free software: you can redistribute it and/or modify
# it under the terms of the GNU LLesser General Public License as
# published by the Free Software Foundation, either version 3 of the License,
# or (at your option) any later version.
#
# SIDEKIT is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU Lesser General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with SIDEKIT.  If not, see <http://www.gnu.org/licenses/>.

"""
Anthony Larcher's avatar
v1.3.7    
Anthony Larcher committed
25
Copyright 2014-2021 Anthony Larcher, Yevhenii Prokopalo
Anthony Larcher's avatar
Anthony Larcher committed
26
"""
Anthony Larcher's avatar
Anthony Larcher committed
27

Anthony Larcher's avatar
Anthony Larcher committed
28

Anthony Larcher's avatar
Anthony Larcher committed
29
import logging
Anthony Larcher's avatar
debug    
Anthony Larcher committed
30
import math
Anthony Larcher's avatar
Anthony Larcher committed
31
import os
Anthony Larcher's avatar
Anthony Larcher committed
32
import numpy
Anthony Larcher's avatar
Anthony Larcher committed
33
import pandas
Anthony Larcher's avatar
minor    
Anthony Larcher committed
34
import pickle
Anthony Larcher's avatar
Anthony Larcher committed
35
import shutil
Anthony Larcher's avatar
Anthony Larcher committed
36
import tabulate
37
import time
Anthony Larcher's avatar
Anthony Larcher committed
38
import torch
Anthony Larcher's avatar
debug    
Anthony Larcher committed
39
import torchaudio
Anthony Larcher's avatar
Anthony Larcher committed
40
import tqdm
Anthony Larcher's avatar
Anthony Larcher committed
41
42
import yaml

Anthony Larcher's avatar
Anthony Larcher committed
43
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
44
45
46
47
48
from torch.utils.data import DataLoader
from sklearn.model_selection import train_test_split
from .preprocessor import MfccFrontEnd
from .preprocessor import MelSpecFrontEnd
from .preprocessor import RawPreprocessor
Anthony Larcher's avatar
Anthony Larcher committed
49
from .xsets import SideSet
Anthony Larcher's avatar
Anthony Larcher committed
50
from .xsets import IdMapSet
Anthony Larcher's avatar
Anthony Larcher committed
51
from .xsets import IdMapSetPerSpeaker
Anthony Larcher's avatar
Anthony Larcher committed
52
from .xsets import SideSampler
Anthony Larcher's avatar
Anthony Larcher committed
53
54
55
56
from .res_net import ResBlockWFMS
from .res_net import ResBlock
from .res_net import PreResNet34
from .res_net import PreFastResNet34
Anthony Larcher's avatar
Anthony Larcher committed
57
from ..bosaris import IdMap
Anthony Larcher's avatar
Anthony Larcher committed
58
59
from ..bosaris import Key
from ..bosaris import Ndx
Anthony Larcher's avatar
Anthony Larcher committed
60
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
61
from ..iv_scoring import cosine_scoring
Anthony Larcher's avatar
Anthony Larcher committed
62
from .sincnet import SincNet
Anthony Larcher's avatar
merge    
Anthony Larcher committed
63
from ..bosaris.detplot import rocch, rocch2eer
Anthony Larcher's avatar
merge    
Anthony Larcher committed
64
from .loss import SoftmaxAngularProto, ArcLinear
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
65
from .loss import l2_norm
Anthony Larcher's avatar
Anthony Larcher committed
66
from .loss import ArcMarginProduct
Anthony Larcher's avatar
Anthony Larcher committed
67

Anthony Larcher's avatar
Anthony Larcher committed
68
from ..sidekit_io import init_logging
Anthony Larcher's avatar
ddp    
Anthony Larcher committed
69

Anthony Larcher's avatar
Anthony Larcher committed
70
os.environ['MKL_THREADING_LAYER'] = 'GNU'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
71

Anthony Larcher's avatar
Anthony Larcher committed
72
73
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
v1.3.7    
Anthony Larcher committed
74
__copyright__ = "Copyright 2015-2021 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
75
76
77
78
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
79
80


Anthony Larcher's avatar
Anthony Larcher committed
81
#logging.basicConfig(format='%(asctime)s %(message)s')
Anthony Larcher's avatar
Anthony Larcher committed
82

Anthony Larcher's avatar
Anthony Larcher committed
83
84
85

# Make PyTorch Deterministic
torch.manual_seed(0)
Anthony Larcher's avatar
Anthony Larcher committed
86
87
torch.backends.cudnn.deterministic = False
torch.backends.cudnn.benchmark = True
Anthony Larcher's avatar
Anthony Larcher committed
88
89
90
numpy.random.seed(0)


Anthony Larcher's avatar
Anthony Larcher committed
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

def eer(negatives, positives):
    """Logarithmic complexity EER computation

    Args:
        negative_scores (numpy array): impostor scores
        positive_scores (numpy array): genuine scores

    Returns:
        float: Equal Error Rate (EER)
    """

    positives = numpy.sort(positives)
    negatives = numpy.sort(negatives)[::-1]

    pos_count = positives.shape[0]
    neg_count = negatives.shape[0]

    p_score = positives[0]
    n_score = negatives[0]

    p_index = 0
    n_index = 0

    next_p_jump = pos_count//2
    next_n_jump = neg_count//2

    kdx = 0
    while True:
        kdx += 1
        if p_index < 0 or n_index < 0:
            return 0
        if p_index > pos_count or n_index > neg_count:
            return 100
        if p_score < n_score:
            p_index = p_index + next_p_jump
            n_index = n_index + next_n_jump
            if next_p_jump == 0 and next_n_jump == 0:
                break
        elif p_score >= n_score:
            p_index = p_index - next_p_jump
            n_index = n_index - next_n_jump
            if next_p_jump == 0 and next_n_jump == 0:
                break

        p_score = positives[p_index]
        n_score = negatives[n_index]
        next_p_jump = next_p_jump//2
        next_n_jump = next_n_jump//2

    eer_predicate = 100

    tfr = (abs(p_index))/pos_count
    tfa = (1+abs(n_index))/neg_count
    if (p_score == n_score and tfr == tfa):
        return tfr

    while positives[p_index] < negatives[n_index]:
        if p_index < pos_count - 1:
            p_index += 1
        elif n_index < neg_count - 1:
            n_index += 1
        else:
            break

    while positives[p_index] > negatives[n_index] and n_index >= 1:
        n_index -= 1

    tfr = (1+p_index)/pos_count
    tfa = (1+n_index)/neg_count

    while tfa > tfr:
        p_index += 1
        while positives[p_index] > negatives[n_index] and n_index >= 1:
            n_index -= 1
        tfr = (1+p_index)/pos_count
        tfa = (1+n_index)/neg_count

    if abs(tfr - tfa) <= eer_predicate:
        eer_predicate = abs(tfr - tfa)
        eer = (tfr + tfa) / 2
    else:
        return eer

    tfr = p_index/pos_count
    tfa = (1+n_index)/neg_count
    if abs(tfr - tfa) <= eer_predicate:
        eer_predicate = abs(tfr - tfa)
        eer = (tfr + tfa) / 2
    else:
        return eer

    while True:
        while negatives[n_index + 1] <= positives[p_index - 1]:
            p_index -= 1
            tfr = p_index/pos_count
            tfa = (1+n_index)/neg_count
            if abs(tfr - tfa) <= eer_predicate:
                eer_predicate = abs(tfr - tfa)
                eer = (tfr + tfa) / 2
            else:
                return eer
        while negatives[n_index + 1] > positives[p_index - 1]:
            n_index += 1
            tfr = p_index/pos_count
            tfa = (1+n_index)/neg_count
            if abs(tfr - tfa) <= eer_predicate:
                eer_predicate = abs(tfr - tfa)
                eer = (tfr + tfa) / 2
            else:
                return eer

    return eer


Anthony Larcher's avatar
debug    
Anthony Larcher committed
206
207
def test_metrics(model,
                 device,
Anthony Larcher's avatar
Anthony Larcher committed
208
                 speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
209
210
                 num_thread,
                 mixed_precision):
Anthony Larcher's avatar
Anthony Larcher committed
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
    """Compute model metrics

    Args:
        model ([type]): [description]
        validation_loader ([type]): [description]
        device ([type]): [description]
        speaker_number ([type]): [description]
        model_archi ([type]): [description]

    Raises:
        NotImplementedError: [description]
        NotImplementedError: [description]

    Returns:
        [type]: [description]
    """
Anthony Larcher's avatar
Anthony Larcher committed
227

Anthony Larcher's avatar
Anthony Larcher committed
228
229
230
231
    idmap_test_filename = '/lium/raid01_c/larcher/data/allies_dev_verif_idmap.h5'
    ndx_test_filename = '/lium/raid01_c/larcher/data/allies_dev_verif_ndx.h5'
    key_test_filename = '/lium/raid01_c/larcher/data/allies_dev_verif_key.h5'
    data_root_name='/lium/corpus/base/ALLIES/wav'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
232

233
    transform_pipeline = dict()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
234
235
236
237
238

    xv_stat = extract_embeddings(idmap_name=idmap_test_filename,
                                 model_filename=model,
                                 data_root_name=data_root_name,
                                 device=device,
Anthony Larcher's avatar
Anthony Larcher committed
239
                                 loss="aam",
Anthony Larcher's avatar
Anthony Larcher committed
240
241
                                 transform_pipeline=transform_pipeline,
                                 num_thread=num_thread,
Anthony Larcher's avatar
Anthony Larcher committed
242
                                 mixed_precision=mixed_precision)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
243

Anthony Larcher's avatar
merge    
Anthony Larcher committed
244
245
246
247
248
249
250
    tar, non = cosine_scoring(xv_stat,
                              xv_stat,
                              Ndx(ndx_test_filename),
                              wccn=None,
                              check_missing=True,
                              device=device
                              ).get_tar_non(Key(key_test_filename))
Anthony Larcher's avatar
debug    
Anthony Larcher committed
251

Anthony Larcher's avatar
merge    
Anthony Larcher committed
252
253
254
255
    #test_eer = eer(numpy.array(non).astype(numpy.double), numpy.array(tar).astype(numpy.double))
    pmiss, pfa = rocch(tar, non)

    return rocch2eer(pmiss, pfa)
Anthony Larcher's avatar
Anthony Larcher committed
256

Anthony Larcher's avatar
Anthony Larcher committed
257
def new_test_metrics(model,
Anthony Larcher's avatar
Anthony Larcher committed
258
259
260
                     device,
                     data_opts,
                     train_opts):
Anthony Larcher's avatar
Anthony Larcher committed
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
    """Compute model metrics

    Args:
        model ([type]): [description]
        validation_loader ([type]): [description]
        device ([type]): [description]
        speaker_number ([type]): [description]
        model_archi ([type]): [description]

    Raises:
        NotImplementedError: [description]
        NotImplementedError: [description]

    Returns:
        [type]: [description]
    """
    # TODO modifier les parametres pour utiliser le dataset_description a la place de :

    #idmap_test_filename,
    #ndx_test_filename,
    #key_test_filename,
    #data_root_name,

    transform_pipeline = dict()

    xv_stat = extract_embeddings(idmap_name=data_opts["idmap_test_filename"],
                                 model_filename=model,
                                 data_root_name=data_opts["data_root_name"],
                                 device=device,
                                 loss=model.loss,
                                 transform_pipeline=transform_pipeline,
                                 num_thread=train_opts["num_thread"],
                                 mixed_precision=train_opts["mixed_precision"])

    tar, non = cosine_scoring(xv_stat,
                              xv_stat,
                              Ndx(data_opts["ndx_test_filename"]),
                              wccn=None,
                              check_missing=True,
                              device=device
                              ).get_tar_non(Key(data_opts["key_test_filename"]))

    #test_eer = eer(numpy.array(non).astype(numpy.double), numpy.array(tar).astype(numpy.double))
    pmiss, pfa = rocch(tar, non)

    return rocch2eer(pmiss, pfa)

Anthony Larcher's avatar
Anthony Larcher committed
308

Anthony Larcher's avatar
Anthony Larcher committed
309
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
310
311
312
313
314
315
316
317
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
318
319
320
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
321

Anthony Larcher's avatar
Anthony Larcher committed
322

Anthony Larcher's avatar
Anthony Larcher committed
323

Anthony Larcher's avatar
Anthony Larcher committed
324
325
326
class TrainingMonitor():

    def __init__(self,
Anthony Larcher's avatar
Anthony Larcher committed
327
                 output_file,
Anthony Larcher's avatar
Anthony Larcher committed
328
                 log_interval=10,
Anthony Larcher's avatar
Anthony Larcher committed
329
330
                 patience=numpy.inf,
                 best_accuracy=0.0,
Anthony Larcher's avatar
Anthony Larcher committed
331
                 best_eer_epoch=1,
Anthony Larcher's avatar
Anthony Larcher committed
332
333
334
335
                 best_eer=100,
                 compute_test_eer=False
                 ):
        # Stocker plutot des listes pour conserver l'historique complet
Anthony Larcher's avatar
Anthony Larcher committed
336
        self.current_epoch = 0
Anthony Larcher's avatar
Anthony Larcher committed
337
        self.log_interval = log_interval
Anthony Larcher's avatar
Anthony Larcher committed
338
339
340
        self.init_patience = patience
        self.current_patience = patience
        self.best_accuracy = best_accuracy
Anthony Larcher's avatar
Anthony Larcher committed
341
        self.best_eer_epoch = best_eer_epoch
Anthony Larcher's avatar
Anthony Larcher committed
342
        self.best_eer = best_eer
Anthony Larcher's avatar
Anthony Larcher committed
343
        self.compute_test_eer = compute_test_eer
Anthony Larcher's avatar
Anthony Larcher committed
344
        self.test_eer = []
Anthony Larcher's avatar
Anthony Larcher committed
345
346
347
348

        self.training_loss = []
        self.training_acc = []

Anthony Larcher's avatar
Anthony Larcher committed
349
350
        self.val_loss = []
        self.val_acc = []
Anthony Larcher's avatar
Anthony Larcher committed
351
352
        self.val_eer = []

Anthony Larcher's avatar
Anthony Larcher committed
353
354
        self.is_best = True

Anthony Larcher's avatar
Anthony Larcher committed
355
        # Initialize the logger
Anthony Larcher's avatar
Anthony Larcher committed
356
        init_logging(level=logging.DEBUG, filename=output_file)
Anthony Larcher's avatar
Anthony Larcher committed
357
358
359
360
361
362
363
        self.logger = logging.getLogger('monitoring')
        self.logger.setLevel(logging.INFO)
        # create file handler which logs even debug messages
        fh = logging.FileHandler(output_file)
        fh.setLevel(logging.INFO)
        self.logger.addHandler(fh)

Anthony Larcher's avatar
Anthony Larcher committed
364
365
366
367
368
369
    def display(self):
        """

        :return:
        """
        # TODO
Anthony Larcher's avatar
Anthony Larcher committed
370
371
        self.logger.critical(f"***{time.strftime('%H:%M:%S', time.localtime())} Training metrics - Cross validation accuracy = {val_acc} %, EER = {val_eer * 100} %")
        self.logger.critical(f"***{time.strftime('%H:%M:%S', time.localtime())} Training metrics - Test EER = {test_eer * 100} %")
Anthony Larcher's avatar
Anthony Larcher committed
372
373
374
375
376
377

    def display_final(self):
        """

        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
378
        self.logger.critical(f"Best accuracy {self.best_accuracy * 100.} obtained at epoch {self.best_accuracy_epoch}")
Anthony Larcher's avatar
Anthony Larcher committed
379
380

    def update(self,
Anthony Larcher's avatar
Anthony Larcher committed
381
               epoch=None,
Anthony Larcher's avatar
Anthony Larcher committed
382
383
               training_acc=None,
               training_loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
384
385
386
387
388
               test_eer=None,
               val_eer=None,
               val_loss=None,
               val_acc=None):

Anthony Larcher's avatar
Anthony Larcher committed
389
390
391
        if epoch is not None:
            self.current_epoch = epoch
        if training_acc is not None:
Anthony Larcher's avatar
Anthony Larcher committed
392
            self.training_acc.append(training_acc)
Anthony Larcher's avatar
Anthony Larcher committed
393
        if training_loss is not None:
Anthony Larcher's avatar
Anthony Larcher committed
394
            self.training_loss.append(training_loss)
Anthony Larcher's avatar
Anthony Larcher committed
395
        if val_eer is not None:
Anthony Larcher's avatar
Anthony Larcher committed
396
            self.val_eer.append(val_eer)
Anthony Larcher's avatar
Anthony Larcher committed
397
        if val_loss is not None:
Anthony Larcher's avatar
Anthony Larcher committed
398
            self.val_loss.append(val_loss)
Anthony Larcher's avatar
Anthony Larcher committed
399
        if val_acc is not None:
Anthony Larcher's avatar
Anthony Larcher committed
400
            self.val_acc.append(val_acc)
Anthony Larcher's avatar
Anthony Larcher committed
401
402

        # remember best accuracy and save checkpoint
Anthony Larcher's avatar
Anthony Larcher committed
403
        if self.compute_test_eer and test_eer is not None:
Anthony Larcher's avatar
Anthony Larcher committed
404
405
406
            self.test_eer.append(test_eer)
            self.is_best = test_eer < self.best_eer
            self.best_eer = min(test_eer, self.best_eer)
Anthony Larcher's avatar
Anthony Larcher committed
407
408
409
410
411
            if self.is_best:
                self.best_eer_epoch = epoch
                self.current_patience = self.init_patience
            else:
                self.current_patience -= 1
Anthony Larcher's avatar
Anthony Larcher committed
412
        elif val_eer is not None:
Anthony Larcher's avatar
Anthony Larcher committed
413
414
            self.is_best = val_eer < self.best_eer
            self.best_eer = min(val_eer, self.best_eer)
Anthony Larcher's avatar
Anthony Larcher committed
415
416
417
418
419
            if self.is_best:
                self.best_eer_epoch = epoch
                self.current_patience = self.init_patience
            else:
                self.current_patience -= 1
Anthony Larcher's avatar
Anthony Larcher committed
420
421


Anthony Larcher's avatar
Anthony Larcher committed
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
class MeanStdPooling(torch.nn.Module):
    """
    Mean and Standard deviation pooling
    """
    def __init__(self):
        """

        """
        super(MeanStdPooling, self).__init__()
        pass

    def forward(self, x):
        """

        :param x:
        :return:
        """
        mean = torch.mean(x, dim=2)
        std = torch.std(x, dim=2)
        return torch.cat([mean, std], dim=1)
Anthony Larcher's avatar
Anthony Larcher committed
442

Anthony Larcher's avatar
Anthony Larcher committed
443

Anthony Larcher's avatar
Anthony Larcher committed
444
445
446
447
448
449
450
451
452
453
454
class GruPooling(torch.nn.Module):
    """

    """
    def __init__(self, input_size, gru_node, nb_gru_layer):
        """

        :param input_size:
        :param gru_node:
        :param nb_gru_layer:
        """
Anthony Larcher's avatar
Anthony Larcher committed
455
        super(GruPooling, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
456
457
458
        self.lrelu_keras = torch.nn.LeakyReLU(negative_slope = 0.3)
        self.bn_before_gru = torch.nn.BatchNorm1d(num_features = input_size)
        self.gru = torch.nn.GRU(input_size = input_size,
Anthony Larcher's avatar
Anthony Larcher committed
459
460
461
                                hidden_size = gru_node,
                                num_layers = nb_gru_layer,
                                batch_first = True)
Anthony Larcher's avatar
Anthony Larcher committed
462
463
464
465
466
467
468
469
470
471
472
473

    def forward(self, x):
        """

        :param x:
        :return:
        """
        x = self.bn_before_gru(x)
        x = self.lrelu_keras(x)
        x = x.permute(0, 2, 1)  #(batch, filt, time) >> (batch, time, filt)
        self.gru.flatten_parameters()
        x, _ = self.gru(x)
Anthony Larcher's avatar
Anthony Larcher committed
474
        x = x[:, -1, :]
Anthony Larcher's avatar
Anthony Larcher committed
475
476
477

        return x

Anthony Larcher's avatar
Anthony Larcher committed
478

Anthony Larcher's avatar
Anthony Larcher committed
479
class Xtractor(torch.nn.Module):
480
481
482
    """
    Class that defines an x-vector extractor based on 5 convolutional layers and a mean standard deviation pooling
    """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
483

Anthony Larcher's avatar
Anthony Larcher committed
484
485
486
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
487
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
488
                 norm_embedding=False,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
489
490
                 aam_margin=0.2,
                 aam_s=30):
Anthony Larcher's avatar
Anthony Larcher committed
491
492
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
493
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
494
        """
Anthony Larcher's avatar
Anthony Larcher committed
495
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
496
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
497
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
498
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
499

Anthony Larcher's avatar
Anthony Larcher committed
500
501
        print(f"Speaker number : {self.speaker_number}")

Anthony Larcher's avatar
Anthony Larcher committed
502
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
503

Anthony Larcher's avatar
Anthony Larcher committed
504
505
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
506
507
508
509
510
            if loss not in ["cce", 'aam']:
                raise NotImplementedError(f"The valid loss are for now cce and aam ")
            else:
                self.loss = loss

Anthony Larcher's avatar
Anthony Larcher committed
511
512
            self.activation = torch.nn.LeakyReLU(0.2)

Anthony Larcher's avatar
Anthony Larcher committed
513
            self.preprocessor = MfccFrontEnd()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
514
            self.feature_size = self.preprocessor.n_mfcc
Anthony Larcher's avatar
Anthony Larcher committed
515

Anthony Larcher's avatar
xv    
Anthony Larcher committed
516
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
517
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
518
                ("activation1", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
519
                ("batch_norm1", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
520
521
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
522
                ("batch_norm2", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
523
524
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
525
                ("batch_norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
526
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
527
                ("activation4", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
528
                ("batch_norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
529
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
530
                ("activation5", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
531
                ("batch_norm5", torch.nn.BatchNorm1d(1536))
Anthony Larcher's avatar
Anthony Larcher committed
532
533
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
534
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
535
            self.stat_pooling_weight_decay = 0
Anthony Larcher's avatar
xv    
Anthony Larcher committed
536
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
537
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
538
539
            ]))

540
541
            self.embedding_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
542
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
543
544
545
546
547
                self.after_speaker_embedding = ArcMarginProduct(512,
                                                                int(self.speaker_number),
                                                                s=64,
                                                                m=0.2,
                                                                easy_margin=True)
Anthony Larcher's avatar
Anthony Larcher committed
548
549
550
            elif self.loss == "cce":
                self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
                    ("activation6", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
551
                    ("batch_norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
552
553
554
                    ("dropout6", torch.nn.Dropout(p=0.05)),
                    ("linear7", torch.nn.Linear(512, 512)),
                    ("activation7", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
555
                    ("batch_norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
556
557
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
558

559
            self.preprocessor_weight_decay = 0.0002
Anthony Larcher's avatar
debug    
Anthony Larcher committed
560
561
562
            self.sequence_network_weight_decay = 0.0002
            self.before_speaker_embedding_weight_decay = 0.002
            self.after_speaker_embedding_weight_decay = 0.002
563
            self.embedding_size = 512
Anthony Larcher's avatar
Anthony Larcher committed
564

Anthony Larcher's avatar
Anthony Larcher committed
565
        elif model_archi == "resnet34":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
566

Anthony Larcher's avatar
Anthony Larcher committed
567
            self.preprocessor = MelSpecFrontEnd(n_mels=80)
Anthony Larcher's avatar
Anthony Larcher committed
568
            self.sequence_network = PreResNet34()
Anthony Larcher's avatar
Anthony Larcher committed
569
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
570

Anthony Larcher's avatar
Anthony Larcher committed
571
572
            self.before_speaker_embedding = torch.nn.Linear(in_features=5120,
                                                            out_features=self.embedding_size)
Anthony Larcher's avatar
Anthony Larcher committed
573
574
575
576

            self.stat_pooling = MeanStdPooling()
            self.stat_pooling_weight_decay = 0

577
            self.loss = "aam"
Anthony Larcher's avatar
Anthony Larcher committed
578
            self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
Anthony Larcher's avatar
Anthony Larcher committed
579
580
                                                            int(self.speaker_number),
                                                            s = 30.0,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
581
                                                            m = 0.20,
Anthony Larcher's avatar
Anthony Larcher committed
582
                                                            easy_margin = False)
Anthony Larcher's avatar
Anthony Larcher committed
583
584
585
586
587
588
589

            self.preprocessor_weight_decay = 0.000
            self.sequence_network_weight_decay = 0.000
            self.stat_pooling_weight_decay = 0.000
            self.before_speaker_embedding_weight_decay = 0.00
            self.after_speaker_embedding_weight_decay = 0.00

Anthony Larcher's avatar
Anthony Larcher committed
590
        elif model_archi == "fastresnet34":
Anthony Larcher's avatar
Anthony Larcher committed
591
            self.preprocessor = MelSpecFrontEnd()
Anthony Larcher's avatar
Anthony Larcher committed
592
            self.sequence_network = PreFastResNet34()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
593
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
594

Anthony Larcher's avatar
Anthony Larcher committed
595
            self.before_speaker_embedding = torch.nn.Linear(in_features = 2560,
Anthony Larcher's avatar
Anthony Larcher committed
596
                                                            out_features = self.embedding_size)
Anthony Larcher's avatar
Anthony Larcher committed
597
598
599
600

            self.stat_pooling = MeanStdPooling()
            self.stat_pooling_weight_decay = 0

Anthony Larcher's avatar
merge    
Anthony Larcher committed
601
602
            self.loss = loss
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
603
                print("\n\n\nAAM\n\n\n")
Anthony Larcher's avatar
merge    
Anthony Larcher committed
604
605
                self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
                                                                int(self.speaker_number),
Anthony Larcher's avatar
merge    
Anthony Larcher committed
606
607
                                                                s = 30,
                                                                m = 0.2,
Anthony Larcher's avatar
merge    
Anthony Larcher committed
608
609
610
611
                                                                easy_margin = False)

            elif self.loss == 'aps':
                self.after_speaker_embedding = SoftmaxAngularProto(int(self.speaker_number))
Anthony Larcher's avatar
Anthony Larcher committed
612
613
614
615
616
617

            self.preprocessor_weight_decay = 0.000
            self.sequence_network_weight_decay = 0.000
            self.stat_pooling_weight_decay = 0.000
            self.before_speaker_embedding_weight_decay = 0.00
            self.after_speaker_embedding_weight_decay = 0.00
Anthony Larcher's avatar
Anthony Larcher committed
618

Anthony Larcher's avatar
Anthony Larcher committed
619
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
620
621
622
623
624
625

            if loss not in ["cce", 'aam']:
                raise NotImplementedError(f"The valid loss are for now cce and aam ")
            else:
                self.loss = loss

Anthony Larcher's avatar
Anthony Larcher committed
626
627
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
628
629
630
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
631
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
632
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
633
634
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651

            self.sequence_network = torch.nn.Sequential(OrderedDict([
                ("block0", ResBlockWFMS(nb_filts=filts[1], first=True)),
                ("block1", ResBlockWFMS(nb_filts=filts[1])),
                ("block2", ResBlockWFMS(nb_filts=filts[2])),
                ("block3", ResBlockWFMS(nb_filts=[filts[2][1], filts[2][1]])),
                ("block4", ResBlockWFMS(nb_filts=[filts[2][1], filts[2][1]])),
                ("block5", ResBlockWFMS(nb_filts=[filts[2][1], filts[2][1]]))
            ]))

            self.stat_pooling = GruPooling(input_size=filts[2][-1],
                                           gru_node=1024,
                                           nb_gru_layer=1)

            self.before_speaker_embedding = torch.nn.Linear(in_features = 1024,
                                                            out_features = 1024)

Anthony Larcher's avatar
Anthony Larcher committed
652
653
654
655
656
            if self.loss == "aam":
                if loss == 'aam':
                    self.after_speaker_embedding = ArcLinear(1024,
                                                             int(self.speaker_number),
                                                             margin=aam_margin, s=aam_s)
Anthony Larcher's avatar
Anthony Larcher committed
657
            elif self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
658
659
660
                self.after_speaker_embedding = torch.nn.Linear(in_features = 1024,
                                                               out_features = int(self.speaker_number),
                                                               bias = True)
Anthony Larcher's avatar
Anthony Larcher committed
661

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
662
663
664
665
666
667
            self.preprocessor_weight_decay = 0.000
            self.sequence_network_weight_decay = 0.000
            self.stat_pooling_weight_decay = 0.000
            self.before_speaker_embedding_weight_decay = 0.00
            self.after_speaker_embedding_weight_decay = 0.00

Anthony Larcher's avatar
Anthony Larcher committed
668
        else:
Anthony Larcher's avatar
Anthony Larcher committed
669
670
            is_first_resblock = True

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
671
672
673
674
675
676
            if isinstance(model_archi, dict):
                cfg = model_archi
            else:
                # Load Yaml configuration
                with open(model_archi, 'r') as fh:
                    cfg = yaml.load(fh, Loader=yaml.FullLoader)
Anthony Larcher's avatar
Anthony Larcher committed
677

Anthony Larcher's avatar
Anthony Larcher committed
678
            self.loss = cfg["loss"]["type"]
Anthony Larcher's avatar
Anthony Larcher committed
679
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
680
681
                self.aam_margin = cfg["loss"]["aam_margin"]
                self.aam_s = cfg["loss"]["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
682

Anthony Larcher's avatar
Anthony Larcher committed
683
684
685
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
686
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
687
688
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
689
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
690
691
692
693
694
695
696
697
698
699
700
701
                        waveform_normalize=cfg['preprocessor']["waveform_normalize"],
                        sample_rate=cfg['preprocessor']["sample_rate"],
                        min_low_hz=cfg['preprocessor']["min_low_hz"],
                        min_band_hz=cfg['preprocessor']["min_band_hz"],
                        out_channels=cfg['preprocessor']["out_channels"],
                        kernel_size=cfg['preprocessor']["kernel_size"],
                        stride=cfg['preprocessor']["stride"],
                        max_pool=cfg['preprocessor']["max_pool"],
                        instance_normalize=cfg['preprocessor']["instance_normalize"],
                        activation=cfg['preprocessor']["activation"],
                        dropout=cfg['preprocessor']["dropout"]
                    )
Anthony Larcher's avatar
Anthony Larcher committed
702
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
703
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
704
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
705
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
706
707
708
709
710
711
                                                        out_channels=cfg["feature_size"],
                                                        kernel_size=cfg['preprocessor']["kernel_size"],
                                                        stride=cfg['preprocessor']["stride"],
                                                        padding=cfg['preprocessor']["padding"],
                                                        dilation=cfg['preprocessor']["dilation"])
                    self.feature_size = cfg["feature_size"]
Anthony Larcher's avatar
Anthony Larcher committed
712
                self.preprocessor_weight_decay = 0.000
Anthony Larcher's avatar
Anthony Larcher committed
713
714

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
715
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
716
            """
Anthony Larcher's avatar
Anthony Larcher committed
717
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
718
            if self.feature_size is None:
Anthony Larcher's avatar
Anthony Larcher committed
719
                self.feature_size = cfg["preprocessor"]["feature_size"]
Anthony Larcher's avatar
Anthony Larcher committed
720

Anthony Larcher's avatar
Anthony Larcher committed
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
            input_size = self.feature_size

            # Get activation function
            if cfg["activation"] == 'LeakyReLU':
                self.activation = torch.nn.LeakyReLU(0.2)
            elif cfg["activation"] == 'PReLU':
                self.activation = torch.nn.PReLU()
            elif cfg["activation"] == 'ReLU6':
                self.activation = torch.nn.ReLU6()
            else:
                self.activation = torch.nn.ReLU()

            # Create sequential object for the first part of the network
            segmental_layers = []
            for k in cfg["segmental"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
736
737
738
739
740
                if k.startswith("lin"):
                    segmental_layers.append((k, torch.nn.Linear(input_size,
                                                                cfg["segmental"][k]["output"])))
                    input_size = cfg["segmental"][k]["output"]

Anthony Larcher's avatar
Anthony Larcher committed
741
742
743
744
745
                elif k.startswith("conv2D"):
                    segmental_layers.append((k, torch.nn.Conv2d(in_channels=1,
                                                                out_channels=entry_conv_out_channels,
                                                                kernel_size=entry_conv_kernel_size,
                                                                padding=3,
Anthony Larcher's avatar
Anthony Larcher committed
746
                                                                stride=1)))
Anthony Larcher's avatar
Anthony Larcher committed
747

Anthony Larcher's avatar
Anthony Larcher committed
748
                elif k.startswith("conv"):
Anthony Larcher's avatar
Anthony Larcher committed
749
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
750
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
751
752
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
753
754
                    input_size = cfg["segmental"][k]["output_channels"]

Anthony Larcher's avatar
Anthony Larcher committed
755
756
                elif k.startswith("ctrans"):
                    segmental_layers.append((k, torch.nn.ConvTranspose1d(input_size,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
757
                                                                         cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
758
759
                                                                         kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                         dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
760
761
762
                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
763
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
Anthony Larcher committed
764
765
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
766
767
768
769
770
771
                elif k.startswith('resblock'):
                    segmental_layers.append((ResBlock(cfg["segmental"][k]["input_channel"],
                                                      cfg["segmental"][k]["output_channel"],
                                                      is_first_resblock)))
                    is_first_resblock = False

Anthony Larcher's avatar
Anthony Larcher committed
772
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
773
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
774

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
775
776
777
778
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
779
            tmp_input_size = input_size * 2
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
780
781
782
783
            if cfg["stat_pooling"]["type"] == "GRU":
                self.stat_pooling = GruPooling(input_size=cfg["stat_pooling"]["input_size"],
                                               gru_node=cfg["stat_pooling"]["gru_node"],
                                               nb_gru_layer=cfg["stat_pooling"]["nb_gru_layer"])
Anthony Larcher's avatar
debug    
Anthony Larcher committed
784
                tmp_input_size = cfg["stat_pooling"]["gru_node"]
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
785

Anthony Larcher's avatar
Anthony Larcher committed
786
787
            self.stat_pooling_weight_decay = cfg["stat_pooling"]["weight_decay"]

Anthony Larcher's avatar
Anthony Larcher committed
788
            """
Anthony Larcher's avatar
Anthony Larcher committed
789
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
790
            """
Anthony Larcher's avatar
Anthony Larcher committed
791
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
792
            input_size = tmp_input_size
Anthony Larcher's avatar
xv    
Anthony Larcher committed
793
794
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
795
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
796
797
                    if cfg["before_embedding"][k]["output"] == "speaker_number":
                        before_embedding_layers.append((k, torch.nn.Linear(input_size, self.speaker_number)))
Anthony Larcher's avatar
Anthony Larcher committed
798
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
799
800
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
801
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
802
803

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
804
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
805

Anthony Larcher's avatar
Anthony Larcher committed
806
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
807
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
808
809

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
810
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
811

Anthony Larcher's avatar
Anthony Larcher committed
812
            self.embedding_size = input_size
Anthony Larcher's avatar
Anthony Larcher committed
813
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
814
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
815

Anthony Larcher's avatar
Anthony Larcher committed
816
            # if loss_criteria is "cce"
Anthony Larcher's avatar
xv    
Anthony Larcher committed
817
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
818
819
820
821
822
823
824
825
            if self.loss == "cce":
                after_embedding_layers = []
                for k in cfg["after_embedding"].keys():
                    if k.startswith("lin"):
                        if cfg["after_embedding"][k]["output"] == "speaker_number":
                            after_embedding_layers.append((k, torch.nn.Linear(input_size, self.speaker_number)))
                        else:
                            after_embedding_layers.append((k, torch.nn.Linear(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
826
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
827
                            input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
828

Anthony Larcher's avatar
Anthony Larcher committed
829
830
831
832
833
                    elif k.startswith('arc'):
                        after_embedding_layers.append((k, ArcLinear(input_size,
                                                                    self.speaker_number,
                                                                    margin=self.aam_margin,
                                                                    s=self.aam_s)))
Anthony Larcher's avatar
Anthony Larcher committed
834

Anthony Larcher's avatar
Anthony Larcher committed
835
836
                    elif k.startswith("activation"):
                        after_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
837

Anthony Larcher's avatar
Anthony Larcher committed
838
839
                    elif k.startswith('batch_norm'):
                        after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
840

Anthony Larcher's avatar
Anthony Larcher committed
841
842
843
844
845
846
                    elif k.startswith('dropout'):
                        after_embedding_layers.append((k, torch.nn.Dropout(p=cfg["after_embedding"][k])))

                self.after_speaker_embedding = torch.nn.Sequential(OrderedDict(after_embedding_layers))

            elif self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
847
848
849
850
851
852
853
                self.norm_embedding = True
                self.after_speaker_embedding = ArcMarginProduct(input_size,
                                                                int(self.speaker_number),
                                                                s=64,
                                                                m=0.2,
                                                                easy_margin=True)

Anthony Larcher's avatar
arcface    
Anthony Larcher committed
854
855
856
857
                #self.after_speaker_embedding = ArcLinear(input_size,
                #                                         self.speaker_number,
                #                                         margin=self.aam_margin,
                #                                         s=self.aam_s)
Anthony Larcher's avatar
Anthony Larcher committed
858
859
860
861
                #self.after_speaker_embedding = ArcFace(embedding_size=input_size,
                #                                       classnum=self.speaker_number,
                #                                       s=64.,
                #                                       m=0.5)
Anthony Larcher's avatar
Anthony Larcher committed
862

Anthony Larcher's avatar
Anthony Larcher committed
863
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
864

Anthony Larcher's avatar
Anthony Larcher committed
865

866
    def forward(self, x, is_eval=False, target=None, extract_after_pooling=False):
867
868
869
        """

        :param x:
870
        :param is_eval: False for training
871
872
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
873
874
875
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
876
        x = self.sequence_network(x)
877

Anthony Larcher's avatar
Anthony Larcher committed
878
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
879
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
880

881
882
883
        if extract_after_pooling:
            return x

Anthony Larcher's avatar
Anthony Larcher committed
884
        x = self.before_speaker_embedding(x)
885

Anthony Larcher's avatar
Anthony Larcher committed
886
        if self.norm_embedding:
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
887
            #x_norm = x.norm(p=2,dim=1, keepdim=True) / 10. # Why  10. ?
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
888
889
890
            #x_norm = torch.linalg.norm(x, ord=2, dim=1, keepdim=True, out=None, dtype=None)
            #x = torch.div(x, x_norm)
            x = l2_norm(x)
Anthony Larcher's avatar
Anthony Larcher committed
891

Anthony Larcher's avatar
Anthony Larcher committed
892
        if self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
893
            if is_eval:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
894
                return x
Anthony Larcher's avatar
debug    
Anthony Larcher committed
895
            else:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
896
                return self.after_speaker_embedding(x), x
Anthony Larcher's avatar
Anthony Larcher committed
897

Anthony Larcher's avatar
merge    
Anthony Larcher committed
898
        elif self.loss in ['aam', 'aps']:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
899
900
            if is_eval:
                x = torch.nn.functional.normalize(x, dim=1)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
901
            else:
Anthony Larcher's avatar
merge    
Anthony Larcher committed
902
                x = self.after_speaker_embedding(x, target=target), torch.nn.functional.normalize(x, dim=1)
Anthony Larcher's avatar
Anthony Larcher committed
903

Anthony Larcher's avatar
Anthony Larcher committed
904
        return x
Anthony Larcher's avatar
Anthony Larcher committed
905

906
907
    def context_size(self):
        context = 1
Anthony Larcher's avatar
Anthony Larcher committed
908
909
910
911
912
913
914
915
916
        if isinstance(self, Xtractor):
            for name, module in self.sequence_network.named_modules():
                if name.startswith("conv"):
                    context += module.dilation[0] * (module.kernel_size[0] - 1)
        else:
            for name, module in self.module.sequence_network.named_modules():
                if name.startswith("conv"):
                    context += module.dilation[0] * (module.kernel_size[0] - 1)
        return context
Anthony Larcher's avatar
Anthony Larcher committed
917

Anthony Larcher's avatar
Anthony Larcher committed
918

Anthony Larcher's avatar
Anthony Larcher committed
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
def fill_dict(target_dict, source_dict, prefix = ""):
    """
    Recursively Fill a dictionary target_dict by taking values from source_dict

    :param target_dict: output dictionary that is initialized with default values
    :param source_dict: input dictionary
    :return:
    """
    for k1, v1 in target_dict.items():

        if isinstance(v1, dict):
            if k1 in source_dict and isinstance(source_dict[k1], dict):
                #print(f"\n{prefix}{k1}")
                fill_dict(v1, source_dict[k1], prefix + "\t")
                #print("\n")
            else:
                #print(f"{prefix}{k1} set to default value: {target_dict[k1]}")
                pass
        else:
            if k1 in source_dict and source_dict[k1] is not None:
                target_dict[k1] = source_dict[k1]
                #print(f"{prefix}{k1} set to: {source_dict[k1]}")
            else:
                #print(f"{prefix}{k1} set to default value: {target_dict[k1]}")
                pass


Anthony Larcher's avatar
Anthony Larcher committed
946
947
def update_training_dictionary(dataset_description,
                               model_description,
Anthony Larcher's avatar
Anthony Larcher committed
948
949
                               training_description,
                               kwargs=None):
Anthony Larcher's avatar
Anthony Larcher committed
950
951
    """

Anthony Larcher's avatar
Anthony Larcher committed
952
953
954
955
    :param dataset_description:
    :param model_description:
    :param training_description:
    :param kwargs:
Anthony Larcher's avatar
Anthony Larcher committed
956
957
958
959
960
961
    :return:
    """
    dataset_opts=dict()
    model_opts=dict()
    training_opts=dict()

Anthony Larcher's avatar
Anthony Larcher committed
962
963
964
965
966
    if isinstance(dataset_description, str) and os.path.isfile(dataset_description):
        with open(dataset_description, 'r') as fh:
            tmp_data_dict = yaml.load(fh, Loader=yaml.FullLoader)
    else:
        tmp_data_dict = dataset_description
Anthony Larcher's avatar
Anthony Larcher committed
967

Anthony Larcher's avatar
Anthony Larcher committed
968
969
970
971
972
    if isinstance(model_description, str) and os.path.isfile(model_description):
        with open(model_description, 'r') as fh:
            tmp_model_dict = yaml.load(fh, Loader=yaml.FullLoader)
    else:
        tmp_model_dict = model_description
Anthony Larcher's avatar
Anthony Larcher committed
973

Anthony Larcher's avatar
Anthony Larcher committed
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
    if isinstance(training_description, str) and os.path.isfile(training_description):
        with open(training_description, 'r') as fh:
            tmp_train_dict = yaml.load(fh, Loader=yaml.FullLoader)
    else:
        tmp_train_dict = training_description

    # Initialize default dictionaries
    dataset_opts["data_path"] = None
    dataset_opts["dataset_csv"] = None
    dataset_opts["data_file_extension"] = ".wav"
    dataset_opts["sample_rate"] = 16000

    dataset_opts["validation_ratio"] = 0.1
    dataset_opts["batch_size"] = 64

    dataset_opts["train"] = dict()
    dataset_opts["train"]["duration"] = 2.
    dataset_opts["train"]["chunk_per_segment"] = -1
    dataset_opts["train"]["overlap"] = 1.9
    dataset_opts["train"]["transformation"] = dict()
    dataset_opts["train"]["transformation"]["pipeline"] = ""
    dataset_opts["train"]["transformation"]["add_noise"] = dict()
    dataset_opts["train"]["transformation"]["add_noise"]["noise_db_csv"] = ""
    dataset_opts["train"]["transformation"]["add_noise"]["data_path"] = ""
    dataset_opts["train"]["transformation"]["add_reverb"] = dict()
    dataset_opts["train"]["transformation"]["add_reverb"]["noise_db_csv"] = ""
    dataset_opts["train"]["transformation"]["add_reverb"]["data_path"] = ""
For faster browsing, not all history is shown. View entire blame