xvector.py 56.9 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
36
import time
Anthony Larcher's avatar
Anthony Larcher committed
37
import torch
Anthony Larcher's avatar
debug    
Anthony Larcher committed
38
import torchaudio
Anthony Larcher's avatar
Anthony Larcher committed
39
import tqdm
Anthony Larcher's avatar
Anthony Larcher committed
40
41
import yaml

Anthony Larcher's avatar
Anthony Larcher committed
42
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
43
44
45
46
47
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
48
from .xsets import SideSet
Anthony Larcher's avatar
Anthony Larcher committed
49
from .xsets import IdMapSet
Anthony Larcher's avatar
Anthony Larcher committed
50
from .xsets import IdMapSetPerSpeaker
Anthony Larcher's avatar
Anthony Larcher committed
51
from .xsets import SideSampler
Anthony Larcher's avatar
Anthony Larcher committed
52
53
54
55
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
56
from ..bosaris import IdMap
Anthony Larcher's avatar
Anthony Larcher committed
57
58
from ..bosaris import Key
from ..bosaris import Ndx
Anthony Larcher's avatar
Anthony Larcher committed
59
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
60
from ..iv_scoring import cosine_scoring
Anthony Larcher's avatar
Anthony Larcher committed
61
from .sincnet import SincNet
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
62
63
from .loss import ArcLinear
from .loss import l2_norm
Anthony Larcher's avatar
Anthony Larcher committed
64
from .loss import ArcMarginProduct
Anthony Larcher's avatar
Anthony Larcher committed
65

Anthony Larcher's avatar
ddp    
Anthony Larcher committed
66

Anthony Larcher's avatar
Anthony Larcher committed
67
os.environ['MKL_THREADING_LAYER'] = 'GNU'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
68

Anthony Larcher's avatar
Anthony Larcher committed
69
70
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
v1.3.7    
Anthony Larcher committed
71
__copyright__ = "Copyright 2015-2021 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
72
73
74
75
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
76
77


Anthony Larcher's avatar
Anthony Larcher committed
78
79
logging.basicConfig(format='%(asctime)s %(message)s')

Anthony Larcher's avatar
Anthony Larcher committed
80
81
82
83
84
85
86
87

# Make PyTorch Deterministic
torch.manual_seed(0)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
numpy.random.seed(0)


Anthony Larcher's avatar
Anthony Larcher committed
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

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
203
204
def test_metrics(model,
                 device,
Anthony Larcher's avatar
Anthony Larcher committed
205
206
207
                 speaker_number,
                 num_thread,
                 mixed_precision):
Anthony Larcher's avatar
Anthony Larcher committed
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
    """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
debug    
Anthony Larcher committed
224
225
226
    idmap_test_filename = 'h5f/idmap_test.h5'
    ndx_test_filename = 'h5f/ndx_test.h5'
    key_test_filename = 'h5f/key_test.h5'
Anthony Larcher's avatar
Anthony Larcher committed
227
    data_root_name='/data/larcher/voxceleb1/test/wav'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
228
229
230
231
232
233
234
235

    transform_pipeline = dict()

    xv_stat = extract_embeddings(idmap_name=idmap_test_filename,
                                 speaker_number=speaker_number,
                                 model_filename=model,
                                 data_root_name=data_root_name,
                                 device=device,
Anthony Larcher's avatar
Anthony Larcher committed
236
237
238
                                 transform_pipeline=transform_pipeline,
                                 num_thread=num_thread,
                                 mixed_precision=mixed_precision)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
239
240
241
242
243
244
245
246
247
248

    scores = cosine_scoring(xv_stat,
                            xv_stat,
                            Ndx(ndx_test_filename),
                            wccn=None,
                            check_missing=True)

    tar, non = scores.get_tar_non(Key(key_test_filename))
    test_eer = eer(numpy.array(non).astype(numpy.double), numpy.array(tar).astype(numpy.double))
    return test_eer
Anthony Larcher's avatar
Anthony Larcher committed
249

Anthony Larcher's avatar
Anthony Larcher committed
250

Anthony Larcher's avatar
Anthony Larcher committed
251
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
252
253
254
255
256
257
258
259
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
260
261
262
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
263

Anthony Larcher's avatar
Anthony Larcher committed
264

Anthony Larcher's avatar
Anthony Larcher committed
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
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
285

Anthony Larcher's avatar
Anthony Larcher committed
286

Anthony Larcher's avatar
Anthony Larcher committed
287
288
289
290
291
292
293
294
295
296
297
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
298
        super(GruPooling, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
299
300
301
        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
302
303
304
                                hidden_size = gru_node,
                                num_layers = nb_gru_layer,
                                batch_first = True)
Anthony Larcher's avatar
Anthony Larcher committed
305
306
307
308
309
310
311
312
313
314
315
316

    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
317
        x = x[:, -1, :]
Anthony Larcher's avatar
Anthony Larcher committed
318
319
320

        return x

Anthony Larcher's avatar
Anthony Larcher committed
321

Anthony Larcher's avatar
Anthony Larcher committed
322
class Xtractor(torch.nn.Module):
323
324
325
    """
    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
326

Anthony Larcher's avatar
Anthony Larcher committed
327
328
329
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
330
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
331
                 norm_embedding=False,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
332
333
                 aam_margin=0.2,
                 aam_s=30):
Anthony Larcher's avatar
Anthony Larcher committed
334
335
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
336
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
337
        """
Anthony Larcher's avatar
Anthony Larcher committed
338
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
339
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
340
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
341
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
342

Anthony Larcher's avatar
Anthony Larcher committed
343
344
        print(f"Speaker number : {self.speaker_number}")

Anthony Larcher's avatar
Anthony Larcher committed
345
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
346

Anthony Larcher's avatar
Anthony Larcher committed
347
348
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
349
350
351
352
353
            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
354
355
            self.activation = torch.nn.LeakyReLU(0.2)

Anthony Larcher's avatar
Anthony Larcher committed
356
            self.preprocessor = MfccFrontEnd()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
357
            self.feature_size = self.preprocessor.n_mfcc
Anthony Larcher's avatar
Anthony Larcher committed
358

Anthony Larcher's avatar
xv    
Anthony Larcher committed
359
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
360
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
361
                ("activation1", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
362
                ("batch_norm1", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
363
364
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
365
                ("batch_norm2", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
366
367
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
368
                ("batch_norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
369
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
370
                ("activation4", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
371
                ("batch_norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
372
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
373
                ("activation5", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
374
                ("batch_norm5", torch.nn.BatchNorm1d(1536))
Anthony Larcher's avatar
Anthony Larcher committed
375
376
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
377
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
378
            self.stat_pooling_weight_decay = 0
Anthony Larcher's avatar
xv    
Anthony Larcher committed
379
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
380
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
381
382
            ]))

383
384
            self.embedding_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
385
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
386
387
388
389
390
                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
391
392
393
            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
394
                    ("batch_norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
395
396
397
                    ("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
398
                    ("batch_norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
399
400
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
401

Anthony Larcher's avatar
debug    
Anthony Larcher committed
402
403
404
            self.sequence_network_weight_decay = 0.0002
            self.before_speaker_embedding_weight_decay = 0.002
            self.after_speaker_embedding_weight_decay = 0.002
405
            self.embedding_size = 512
Anthony Larcher's avatar
Anthony Larcher committed
406

Anthony Larcher's avatar
Anthony Larcher committed
407
        elif model_archi == "resnet34":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
408

Anthony Larcher's avatar
Anthony Larcher committed
409
            self.preprocessor = MfccFrontEnd()
Anthony Larcher's avatar
Anthony Larcher committed
410
411
412
413
414
415
416
417
            self.sequence_network = PreResNet34()

            self.before_speaker_embedding = torch.nn.Linear(in_features = 5120,
                                                            out_features = 256)

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

418
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
419

420
            self.loss = "aam"
Anthony Larcher's avatar
Anthony Larcher committed
421
422
423
            self.after_speaker_embedding = ArcMarginProduct(256,
                                                            int(self.speaker_number),
                                                            s = 30.0,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
424
                                                            m = 0.20,
Anthony Larcher's avatar
Anthony Larcher committed
425
                                                            easy_margin = True)
Anthony Larcher's avatar
Anthony Larcher committed
426
427
428
429
430
431
432

            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
433
        elif model_archi == "fastresnet34":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
434

Anthony Larcher's avatar
FB    
Anthony Larcher committed
435
            self.preprocessor = MelSpecFrontEnd()
Anthony Larcher's avatar
Anthony Larcher committed
436
            self.sequence_network = PreFastResNet34()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
437
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
438

Anthony Larcher's avatar
Anthony Larcher committed
439
            self.before_speaker_embedding = torch.nn.Linear(in_features = 2560,
Anthony Larcher's avatar
Anthony Larcher committed
440
                                                            out_features = self.embedding_size)
Anthony Larcher's avatar
Anthony Larcher committed
441
442
443
444
445

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

            self.loss = "aam"
Anthony Larcher's avatar
debug    
Anthony Larcher committed
446
            self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
Anthony Larcher's avatar
Anthony Larcher committed
447
                                                            int(self.speaker_number),
Anthony Larcher's avatar
debug    
Anthony Larcher committed
448
449
                                                            s = 30,
                                                            m = 0.2,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
450
                                                            easy_margin = False)
Anthony Larcher's avatar
Anthony Larcher committed
451
452
453
454
455
456

            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
457

Anthony Larcher's avatar
Anthony Larcher committed
458
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
459
460
461
462
463
464

            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
465
466
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
467
468
469
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
470
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
471
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
472
473
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490

            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
491
492
493
494
495
            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
496
            elif self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
497
498
499
                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
500

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
501
502
503
504
505
506
            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
507
        else:
Anthony Larcher's avatar
Anthony Larcher committed
508
509
            is_first_resblock = True

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
510
511
512
513
514
515
            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
516

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
517
            self.loss = cfg["training"]["loss"]
Anthony Larcher's avatar
Anthony Larcher committed
518
            if self.loss == "aam":
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
519
520
                self.aam_margin = cfg["training"]["aam_margin"]
                self.aam_s = cfg["training"]["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
521

Anthony Larcher's avatar
Anthony Larcher committed
522
523
524
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
525
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
526
527
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
528
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
529
530
531
532
533
534
535
536
537
538
539
540
                        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
541
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
542
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
543
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
544
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
545
546
547
548
549
550
                                                        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
551
                self.preprocessor_weight_decay = 0.000
Anthony Larcher's avatar
Anthony Larcher committed
552
553

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
554
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
555
            """
Anthony Larcher's avatar
Anthony Larcher committed
556
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
557
558
559
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
            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
575
576
577
578
579
                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
580
581
582
583
584
                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
585
                                                                stride=1)))
Anthony Larcher's avatar
Anthony Larcher committed
586

Anthony Larcher's avatar
Anthony Larcher committed
587
                elif k.startswith("conv"):
Anthony Larcher's avatar
Anthony Larcher committed
588
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
589
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
590
591
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
592
593
                    input_size = cfg["segmental"][k]["output_channels"]

Anthony Larcher's avatar
Anthony Larcher committed
594
595
                elif k.startswith("ctrans"):
                    segmental_layers.append((k, torch.nn.ConvTranspose1d(input_size,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
596
                                                                         cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
597
598
                                                                         kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                         dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
599
600
601
                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
602
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
Anthony Larcher committed
603
604
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
605
606
607
608
609
610
                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
611
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
612
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
613

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
614
615
616
617
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
618
            tmp_input_size = input_size * 2
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
619
620
621
622
            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
623
                tmp_input_size = cfg["stat_pooling"]["gru_node"]
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
624

Anthony Larcher's avatar
Anthony Larcher committed
625
626
            self.stat_pooling_weight_decay = cfg["stat_pooling"]["weight_decay"]

Anthony Larcher's avatar
Anthony Larcher committed
627
            """
Anthony Larcher's avatar
Anthony Larcher committed
628
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
629
            """
Anthony Larcher's avatar
Anthony Larcher committed
630
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
631
            input_size = tmp_input_size
Anthony Larcher's avatar
xv    
Anthony Larcher committed
632
633
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
634
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
635
636
                    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
637
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
638
639
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
640
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
641
642

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
643
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
644

Anthony Larcher's avatar
Anthony Larcher committed
645
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
646
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
647
648

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
649
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
650

Anthony Larcher's avatar
Anthony Larcher committed
651
            self.embedding_size = input_size
Anthony Larcher's avatar
Anthony Larcher committed
652
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
653
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
654

Anthony Larcher's avatar
Anthony Larcher committed
655
            # if loss_criteria is "cce"
Anthony Larcher's avatar
xv    
Anthony Larcher committed
656
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
657
658
659
660
661
662
663
664
            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
665
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
666
                            input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
667

Anthony Larcher's avatar
Anthony Larcher committed
668
669
670
671
672
                    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
673

Anthony Larcher's avatar
Anthony Larcher committed
674
675
                    elif k.startswith("activation"):
                        after_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
676

Anthony Larcher's avatar
Anthony Larcher committed
677
678
                    elif k.startswith('batch_norm'):
                        after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
679

Anthony Larcher's avatar
Anthony Larcher committed
680
681
682
683
684
685
                    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
686
687
688
689
690
691
692
                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
693
694
695
696
                #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
697
698
699
700
                #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
701

Anthony Larcher's avatar
Anthony Larcher committed
702
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
703

Anthony Larcher's avatar
Anthony Larcher committed
704

705
    def forward(self, x, is_eval=False, target=None, extract_after_pooling=False):
706
707
708
        """

        :param x:
709
        :param is_eval: False for training
710
711
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
712
713
714
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
715
        x = self.sequence_network(x)
716

Anthony Larcher's avatar
Anthony Larcher committed
717
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
718
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
719

720
721
722
        if extract_after_pooling:
            return x

Anthony Larcher's avatar
Anthony Larcher committed
723
        x = self.before_speaker_embedding(x)
724

Anthony Larcher's avatar
Anthony Larcher committed
725
        if self.norm_embedding:
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
726
            #x_norm = x.norm(p=2,dim=1, keepdim=True) / 10. # Why  10. ?
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
727
728
729
            #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
730

Anthony Larcher's avatar
Anthony Larcher committed
731
        if self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
732
            if is_eval:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
733
                return x
Anthony Larcher's avatar
debug    
Anthony Larcher committed
734
            else:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
735
                return self.after_speaker_embedding(x), x
Anthony Larcher's avatar
Anthony Larcher committed
736

Anthony Larcher's avatar
Anthony Larcher committed
737
        elif self.loss == "aam":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
738
739
            if is_eval:
                x = torch.nn.functional.normalize(x, dim=1)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
740
            else:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
741
                x = self.after_speaker_embedding(torch.nn.functional.normalize(x, dim=1), target=target), torch.nn.functional.normalize(x, dim=1)
Anthony Larcher's avatar
Anthony Larcher committed
742

Anthony Larcher's avatar
Anthony Larcher committed
743
        return x
Anthony Larcher's avatar
Anthony Larcher committed
744

745
746
    def context_size(self):
        context = 1
Anthony Larcher's avatar
Anthony Larcher committed
747
748
749
750
751
752
753
754
755
        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
756

Anthony Larcher's avatar
Anthony Larcher committed
757
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
758
           dataset_yaml,
Anthony Larcher's avatar
Anthony Larcher committed
759
760
           epochs=None,
           lr=None,
Anthony Larcher's avatar
Anthony Larcher committed
761
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
762
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
763
764
765
766
           loss=None,
           aam_margin=None,
           aam_s=None,
           patience=None,
Anthony Larcher's avatar
Anthony Larcher committed
767
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
768
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
769
           multi_gpu=True,
Anthony Larcher's avatar
Anthony Larcher committed
770
           mixed_precision=False,
771
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
772
           opt=None,
Anthony Larcher's avatar
Anthony Larcher committed
773
774
           reset_parts=[],
           freeze_parts=[],
Anthony Larcher's avatar
Anthony Larcher committed
775
           num_thread=None,
Anthony Larcher's avatar
Anthony Larcher committed
776
           compute_test_eer=True):
777
778
    """

Anthony Larcher's avatar
Anthony Larcher committed
779
780
781
782
783
784
    :param speaker_number:
    :param dataset_yaml:
    :param epochs:
    :param lr:
    :param model_yaml:
    :param model_name:
Anthony Larcher's avatar
Anthony Larcher committed
785
786
787
788
    :param loss:
    :param aam_margin:
    :param aam_s:
    :param patience:
Anthony Larcher's avatar
Anthony Larcher committed
789
790
791
    :param tmp_model_name:
    :param best_model_name:
    :param multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
792
    :param mixed_precision:
Anthony Larcher's avatar
Anthony Larcher committed
793
    :param clipping:
Anthony Larcher's avatar
Anthony Larcher committed
794
795
796
    :param opt:
    :param reset_parts:
    :param freeze_parts:
Anthony Larcher's avatar
Anthony Larcher committed
797
    :param num_thread:
Anthony Larcher's avatar
Anthony Larcher committed
798
    :param compute_test_eer:
799
800
    :return:
    """
Anthony Larcher's avatar
debug    
Anthony Larcher committed
801
802
803
    # Test to optimize
    torch.autograd.profiler.emit_nvtx(enabled=False)

Anthony Larcher's avatar
Anthony Larcher committed
804
    if num_thread is None:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
805
        import multiprocessing
Anthony Larcher's avatar
Anthony Larcher committed
806
807
        num_thread = multiprocessing.cpu_count()

Anthony Larcher's avatar
Anthony Larcher committed
808
    logging.critical(f"Use {num_thread} cpus")
Anthony Larcher's avatar
Anthony Larcher committed
809
    logging.critical(f"Start process at {time.strftime('%H:%M:%S', time.localtime())}")
810

Anthony Larcher's avatar
debug    
Anthony Larcher committed
811
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840

    # Use a predefined architecture
    if model_yaml in ["xvector", "rawnet2", "resnet34", "fastresnet34"]:

        if model_name is None:
            model = Xtractor(speaker_number, model_yaml)

        else:
            logging.critical(f"*** Load model from = {model_name}")
            checkpoint = torch.load(model_name)
            model = Xtractor(speaker_number, model_yaml)

            """
            Here we remove all layers that we don't want to reload

            """
            pretrained_dict = checkpoint["model_state_dict"]
            for part in reset_parts:
                pretrained_dict = {k: v for k, v in pretrained_dict.items() if not k.startswith(part)}

            new_model_dict = model.state_dict()
            new_model_dict.update(pretrained_dict)
            model.load_state_dict(new_model_dict)

        # Freeze required layers
        for name, param in model.named_parameters():
            if name.split(".")[0] in freeze_parts:
                param.requires_grad = False

Anthony Larcher's avatar
Anthony Larcher committed
841
        model_archi = model_yaml
842
843

    # Here use a config file to build the architecture
Anthony Larcher's avatar
Anthony Larcher committed
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
    else:
        with open(model_yaml, 'r') as fh:
            model_archi = yaml.load(fh, Loader=yaml.FullLoader)
            if epochs is None:
                epochs = model_archi["training"]["epochs"]
            if patience is None:
                patience = model_archi["training"]["patience"]
            if opt is None:
                opt = model_archi["training"]["opt"]
            if lr is None:
                lr = model_archi["training"]["lr"]
            if loss is None:
                loss = model_archi["training"]["loss"]
            if aam_margin is None and model_archi["training"]["loss"] == "aam":
                aam_margin = model_archi["training"]["aam_margin"]
            if aam_s is None and model_archi["training"]["loss"] == "aam":
                aam_s = model_archi["training"]["aam_s"]
            if tmp_model_name is None:
                tmp_model_name = model_archi["training"]["tmp_model_name"]
            if best_model_name is None:
                best_model_name = model_archi["training"]["best_model_name"]
            if multi_gpu is None:
                multi_gpu = model_archi["training"]["multi_gpu"]
            if clipping is None:
                clipping = model_archi["training"]["clipping"]

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
870
        if model_name is None:
Anthony Larcher's avatar
Anthony Larcher committed
871
872
873
            model = Xtractor(speaker_number, model_yaml)

         # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
874
        else:
Anthony Larcher's avatar
Anthony Larcher committed
875
876
877
            # Load the model
            logging.critical(f"*** Load model from = {model_name}")
            checkpoint = torch.load(model_name)
Anthony Larcher's avatar
Anthony Larcher committed
878
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
879

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
880
881
            """
            Here we remove all layers that we don't want to reload
Anthony Larcher's avatar
Anthony Larcher committed
882
        
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
883
884
885
886
887
888
889
890
891
892
893
894
895
            """
            pretrained_dict = checkpoint["model_state_dict"]
            for part in reset_parts:
                pretrained_dict = {k: v for k, v in pretrained_dict.items() if not k.startswith(part)}

            new_model_dict = model.state_dict()
            new_model_dict.update(pretrained_dict)
            model.load_state_dict(new_model_dict)

        # Freeze required layers
        for name, param in model.named_parameters():
            if name.split(".")[0] in freeze_parts:
                param.requires_grad = False
Anthony Larcher's avatar
Anthony Larcher committed
896

Anthony Larcher's avatar
Anthony Larcher committed
897
898
899
900
    logging.critical(model)

    logging.critical("model_parameters_count: {:d}".format(
        sum(p.numel()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
901
902
903
904
            for p in model.sequence_network.parameters()
            if p.requires_grad) + \
        sum(p.numel()
            for p in model.before_speaker_embedding.parameters()
Anthony Larcher's avatar
Anthony Larcher committed
905
906
            if p.requires_grad)))

Anthony Larcher's avatar
Anthony Larcher committed
907
    embedding_size = model.embedding_size
Anthony Larcher's avatar
Anthony Larcher committed
908

Anthony Larcher's avatar
Anthony Larcher committed
909
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
910
911
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
912

Anthony Larcher's avatar
Anthony Larcher committed
913
914
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
915
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
916

Anthony Larcher's avatar
debug    
Anthony Larcher committed
917
918
919
920
    with open(dataset_yaml, "r") as fh:
        dataset_params = yaml.load(fh, Loader=yaml.FullLoader)
        df = pandas.read_csv(dataset_params["dataset_description"])

Anthony Larcher's avatar
Anthony Larcher committed
921
922
923
924
925
926
927
    """
    Set the dataloaders according to the dataset_yaml
    
    First we load the dataframe from CSV file in order to split it for training and validation purpose
    Then we provide those two 
    """
    training_df, validation_df = train_test_split(df, test_size=dataset_params["validation_ratio"], stratify=df["speaker_idx"])
928

Anthony Larcher's avatar
Anthony Larcher committed
929
    torch.manual_seed(dataset_params['seed'])
Anthony Larcher's avatar
Anthony Larcher committed
930

Anthony Larcher's avatar
Anthony Larcher committed
931
932
933
934
935
936
937
938
939
940
941
942
    training_set = SideSet(dataset_yaml,
                           set_type="train",
                           chunk_per_segment=-1,
                           overlap=dataset_params['train']['overlap'],
                           dataset_df=training_df,
                           output_format="pytorch",
                           )

    validation_set = SideSet(dataset_yaml,
                             set_type="validation",
                             dataset_df=validation_df,
                             output_format="pytorch")
Anthony Larcher's avatar
debug    
Anthony Larcher committed
943

Anthony Larcher's avatar
Anthony Larcher committed
944
945
946
947
    side_sampler = SideSampler(training_set.sessions['speaker_idx'],
                               speaker_number,
                               1,
                               100,
Anthony Larcher's avatar
Anthony Larcher committed
948
                               dataset_params["batch_size"])
Anthony Larcher's avatar
Anthony Larcher committed
949

Anthony Larcher's avatar
debug    
Anthony Larcher committed
950
    training_loader = DataLoader(training_set,
Anthony Larcher's avatar
Anthony Larcher committed
951
                                 batch_size=dataset_params["batch_size"],
Anthony Larcher's avatar
Anthony Larcher committed
952
                                 shuffle=False,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
953
954
                                 drop_last=True,
                                 pin_memory=True,
Anthony Larcher's avatar
Anthony Larcher committed
955
                                 sampler=side_sampler,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
956
                                 num_workers=num_thread,
Anthony Larcher's avatar
Anthony Larcher committed
957
                                 persistent_workers=True)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
958
959

    validation_loader = DataLoader(validation_set,
Anthony Larcher's avatar
Anthony Larcher committed
960
                                   batch_size=dataset_params["batch_size"],
Anthony Larcher's avatar
Anthony Larcher committed
961
                                   drop_last=False,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
962
                                   pin_memory=True,
Anthony Larcher's avatar
spkset    
Anthony Larcher committed
963
                                   num_workers=num_thread,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
964
                                   persistent_workers=False)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
965

Anthony Larcher's avatar
Anthony Larcher committed
966
967
968
    """
    Set the training options
    """
Anthony Larcher's avatar
Anthony Larcher committed
969
    if opt == 'adam':
Anthony Larcher's avatar
Anthony Larcher committed
970
        _optimizer = torch.optim.Adam
Anthony Larcher's avatar
Anthony Larcher committed
971
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
972
973
    elif opt == 'rmsprop':
        _optimizer = torch.optim.RMSprop
Anthony Larcher's avatar
Anthony Larcher committed
974
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
975
976
977
    else: # opt == 'sgd'
        _optimizer = torch.optim.SGD
        _options = {'lr': lr, 'momentum': 0.9}
Anthony Larcher's avatar
Anthony Larcher committed
978

Anthony Larcher's avatar
Anthony Larcher committed
979
    param_list = []
Anthony Larcher's avatar
Anthony Larcher committed
980
    if type(model) is Xtractor:
Anthony Larcher's avatar
Anthony Larcher committed
981
982
983
984
985
986
987
        if model.preprocessor is not None:
            param_list.append({'params': model.preprocessor.parameters(), 'weight_decay': model.preprocessor_weight_decay})
        param_list.append({'params': model.sequence_network.parameters(), 'weight_decay': model.sequence_network_weight_decay})
        param_list.append({'params': model.stat_pooling.parameters(), 'weight_decay': model.stat_pooling_weight_decay})
        param_list.append({'params': model.before_speaker_embedding.parameters(), 'weight_decay': model.before_speaker_embedding_weight_decay})
        param_list.append({'params': model.after_speaker_embedding.parameters(), 'weight_decay': model.after_speaker_embedding_weight_decay})

Anthony Larcher's avatar
Anthony Larcher committed
988
    else:
Anthony Larcher's avatar
Anthony Larcher committed
989
990
991
992
993
994
995
996
        if model.module.preprocessor is not None:
            param_list.append({'params': model.module.preprocessor.parameters(), 'weight_decay': model.module.preprocessor_weight_decay})
        param_list.append({'params': model.module.sequence_network.parameters(), 'weight_decay': model.module.sequence_network_weight_decay})
        param_list.append({'params': model.module.stat_pooling.parameters(), 'weight_decay': model.module.stat_pooling_weight_decay})
        param_list.append({'params': model.module.before_speaker_embedding.parameters(), 'weight_decay': model.module.before_speaker_embedding_weight_decay})
        param_list.append({'params': model.module.after_speaker_embedding.parameters(), 'weight_decay': model.module.after_speaker_embedding_weight_decay})

    optimizer = _optimizer(param_list, **_options)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
997
998
999
    scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
                                                step_size=10 * training_loader.__len__(),
                                                gamma=0.95)
1000