xvector.py 73.3 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 random
Anthony Larcher's avatar
Anthony Larcher committed
34
import pandas
Anthony Larcher's avatar
Anthony Larcher committed
35
import shutil
Anthony Larcher's avatar
Anthony Larcher committed
36
import torch
Anthony Larcher's avatar
Anthony Larcher committed
37
import tqdm
Anthony Larcher's avatar
Anthony Larcher committed
38
39
import yaml

Anthony Larcher's avatar
Anthony Larcher committed
40
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
41
42
from torch.utils.data import DataLoader
from sklearn.model_selection import train_test_split
Anthony Larcher's avatar
Anthony Larcher committed
43
44
45
from .pooling import MeanStdPooling
from .pooling import AttentivePooling
from .pooling import GruPooling
Anthony Larcher's avatar
Anthony Larcher committed
46
47
48
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
from .res_net import ResBlockWFMS
from .res_net import ResBlock
from .res_net import PreFastResNet34
Anthony Larcher's avatar
Anthony Larcher committed
56
57
from .res_net import PreHalfResNet34
from .res_net import PreResNet34
Anthony Larcher's avatar
Anthony Larcher committed
58
from ..bosaris import IdMap
Anthony Larcher's avatar
Anthony Larcher committed
59
60
from ..bosaris import Key
from ..bosaris import Ndx
Anthony Larcher's avatar
Anthony Larcher committed
61
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
62
from ..iv_scoring import cosine_scoring
Anthony Larcher's avatar
Anthony Larcher committed
63
from .sincnet import SincNet
Anthony Larcher's avatar
merge    
Anthony Larcher committed
64
from ..bosaris.detplot import rocch, rocch2eer
Anthony Larcher's avatar
Anthony Larcher committed
65
from .loss import SoftmaxAngularProto
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
66
from .loss import l2_norm
Anthony Larcher's avatar
Anthony Larcher committed
67
from .loss import ArcMarginProduct
Anthony Larcher's avatar
Anthony Larcher committed
68
69
from .loss import ArcLinear
from .loss import AngularProximityMagnet
Anthony Larcher's avatar
Anthony Larcher committed
70

Anthony Larcher's avatar
Anthony Larcher committed
71

Anthony Larcher's avatar
Anthony Larcher committed
72
os.environ['MKL_THREADING_LAYER'] = 'GNU'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
73

Anthony Larcher's avatar
Anthony Larcher committed
74
75
#torch.backends.cudnn.benchmark = True

Anthony Larcher's avatar
Anthony Larcher committed
76
77
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
v1.3.7    
Anthony Larcher committed
78
__copyright__ = "Copyright 2015-2021 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
79
80
81
82
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
83
84


Anthony Larcher's avatar
debug    
Anthony Larcher committed
85
def seed_worker(seed_val):
Anthony Larcher's avatar
Anthony Larcher committed
86
    """
Anthony Larcher's avatar
Anthony Larcher committed
87

Anthony Larcher's avatar
Anthony Larcher committed
88
89
90
91
92
93
    :param worker_id:
    :return:
    """
    worker_seed = torch.initial_seed() % 2**32
    numpy.random.seed(worker_seed)
    random.seed(worker_seed)
Anthony Larcher's avatar
Anthony Larcher committed
94

Anthony Larcher's avatar
Anthony Larcher committed
95
96

def eer(negatives, positives):
Anthony Larcher's avatar
Anthony Larcher committed
97
    """
Anthony Larcher's avatar
Anthony Larcher committed
98
    Logarithmic complexity EER computation
Anthony Larcher's avatar
Anthony Larcher committed
99

Anthony Larcher's avatar
Anthony Larcher committed
100
101
102
103
    :param negatives: negative_scores (numpy array): impostor scores
    :param positives: positive_scores (numpy array): genuine scores
    :return: float: Equal Error Rate (EER)
    """
Anthony Larcher's avatar
Anthony Larcher committed
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
    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
207
208
def test_metrics(model,
                 device,
Anthony Larcher's avatar
Anthony Larcher committed
209
210
211
                 model_opts,
                 data_opts,
                 train_opts):
Anthony Larcher's avatar
Anthony Larcher committed
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
    """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]
    """
    transform_pipeline = dict()

Anthony Larcher's avatar
Anthony Larcher committed
230
    xv_stat = extract_embeddings(idmap_name=data_opts["test"]["idmap"],
Anthony Larcher's avatar
Anthony Larcher committed
231
                                 model_filename=model,
Anthony Larcher's avatar
Anthony Larcher committed
232
                                 data_root_name=data_opts["test"]["data_path"],
Anthony Larcher's avatar
Anthony Larcher committed
233
234
                                 device=device,
                                 transform_pipeline=transform_pipeline,
Anthony Larcher's avatar
Anthony Larcher committed
235
                                 num_thread=train_opts["num_cpu"],
Anthony Larcher's avatar
Anthony Larcher committed
236
237
238
239
                                 mixed_precision=train_opts["mixed_precision"])

    tar, non = cosine_scoring(xv_stat,
                              xv_stat,
Anthony Larcher's avatar
Anthony Larcher committed
240
                              Ndx(data_opts["test"]["ndx"]),
Anthony Larcher's avatar
Anthony Larcher committed
241
242
243
                              wccn=None,
                              check_missing=True,
                              device=device
Anthony Larcher's avatar
Anthony Larcher committed
244
                              ).get_tar_non(Key(data_opts["test"]["key"]))
Anthony Larcher's avatar
Anthony Larcher committed
245
246
247
248
249

    pmiss, pfa = rocch(tar, non)

    return rocch2eer(pmiss, pfa)

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
class TrainingMonitor():
Anthony Larcher's avatar
Anthony Larcher committed
266
    """
Anthony Larcher's avatar
Anthony Larcher committed
267

Anthony Larcher's avatar
Anthony Larcher committed
268
    """
Anthony Larcher's avatar
Anthony Larcher committed
269
    def __init__(self,
Anthony Larcher's avatar
Anthony Larcher committed
270
                 output_file,
Anthony Larcher's avatar
Anthony Larcher committed
271
                 log_interval=10,
Anthony Larcher's avatar
Anthony Larcher committed
272
273
                 patience=numpy.inf,
                 best_accuracy=0.0,
Anthony Larcher's avatar
Anthony Larcher committed
274
                 best_eer_epoch=1,
Anthony Larcher's avatar
Anthony Larcher committed
275
276
277
                 best_eer=100,
                 compute_test_eer=False
                 ):
Anthony Larcher's avatar
Anthony Larcher committed
278

Anthony Larcher's avatar
Anthony Larcher committed
279
        self.current_epoch = 0
Anthony Larcher's avatar
Anthony Larcher committed
280
        self.log_interval = log_interval
Anthony Larcher's avatar
Anthony Larcher committed
281
282
283
        self.init_patience = patience
        self.current_patience = patience
        self.best_accuracy = best_accuracy
Anthony Larcher's avatar
Anthony Larcher committed
284
        self.best_eer_epoch = best_eer_epoch
Anthony Larcher's avatar
Anthony Larcher committed
285
        self.best_eer = best_eer
Anthony Larcher's avatar
Anthony Larcher committed
286
        self.compute_test_eer = compute_test_eer
Anthony Larcher's avatar
Anthony Larcher committed
287
        self.test_eer = []
Anthony Larcher's avatar
Anthony Larcher committed
288
289
290
291

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

Anthony Larcher's avatar
Anthony Larcher committed
292
293
        self.val_loss = []
        self.val_acc = []
Anthony Larcher's avatar
Anthony Larcher committed
294
295
        self.val_eer = []

Anthony Larcher's avatar
Anthony Larcher committed
296
297
        self.is_best = True

Anthony Larcher's avatar
Anthony Larcher committed
298
        # Initialize the logger
Anthony Larcher's avatar
Anthony Larcher committed
299
        logging_format = '%(asctime)-15s %(message)s'
Anthony Larcher's avatar
Anthony Larcher committed
300
        logging.basicConfig(level=logging.DEBUG, format=logging_format, datefmt='%m-%d %H:%M')
Anthony Larcher's avatar
Anthony Larcher committed
301
302
        self.logger = logging.getLogger('Monitoring')
        self.logger.setLevel(logging.DEBUG)
Anthony Larcher's avatar
Anthony Larcher committed
303
304
        # create file handler which logs even debug messages
        fh = logging.FileHandler(output_file)
Anthony Larcher's avatar
Anthony Larcher committed
305
306
        formatter = logging.Formatter('%(asctime)s - %(levelname)s - %(message)s')
        fh.setFormatter(formatter)
Anthony Larcher's avatar
Anthony Larcher committed
307
        fh.setLevel(logging.DEBUG)
Anthony Larcher's avatar
Anthony Larcher committed
308
        self.logger.addHandler(fh)
Anthony Larcher's avatar
Anthony Larcher committed
309

Anthony Larcher's avatar
Anthony Larcher committed
310
311
312
313
314
315
    def display(self):
        """

        :return:
        """
        # TODO
Anthony Larcher's avatar
Anthony Larcher committed
316
317
        self.logger.critical(f"***Validation metrics - Cross validation accuracy = {self.val_acc[-1]} %, EER = {self.val_eer[-1] * 100} %")
        self.logger.critical(f"***Test metrics - Test EER = {self.test_eer[-1] * 100} %")
Anthony Larcher's avatar
Anthony Larcher committed
318
319
320
321
322
323

    def display_final(self):
        """

        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
324
        self.logger.critical(f"Best accuracy {self.best_accuracy * 100.} obtained at epoch {self.best_accuracy_epoch}")
Anthony Larcher's avatar
Anthony Larcher committed
325
326

    def update(self,
Anthony Larcher's avatar
Anthony Larcher committed
327
               epoch=None,
Anthony Larcher's avatar
Anthony Larcher committed
328
329
               training_acc=None,
               training_loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
330
331
332
333
334
               test_eer=None,
               val_eer=None,
               val_loss=None,
               val_acc=None):

Anthony Larcher's avatar
Anthony Larcher committed
335
336
337
        if epoch is not None:
            self.current_epoch = epoch
        if training_acc is not None:
Anthony Larcher's avatar
Anthony Larcher committed
338
            self.training_acc.append(training_acc)
Anthony Larcher's avatar
Anthony Larcher committed
339
        if training_loss is not None:
Anthony Larcher's avatar
Anthony Larcher committed
340
            self.training_loss.append(training_loss)
Anthony Larcher's avatar
Anthony Larcher committed
341
        if val_eer is not None:
Anthony Larcher's avatar
Anthony Larcher committed
342
            self.val_eer.append(val_eer)
Anthony Larcher's avatar
Anthony Larcher committed
343
        if val_loss is not None:
Anthony Larcher's avatar
Anthony Larcher committed
344
            self.val_loss.append(val_loss)
Anthony Larcher's avatar
Anthony Larcher committed
345
        if val_acc is not None:
Anthony Larcher's avatar
Anthony Larcher committed
346
            self.val_acc.append(val_acc)
Anthony Larcher's avatar
Anthony Larcher committed
347
348

        # remember best accuracy and save checkpoint
Anthony Larcher's avatar
Anthony Larcher committed
349
        if self.compute_test_eer and test_eer is not None:
Anthony Larcher's avatar
Anthony Larcher committed
350
351
352
            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
353
354
355
356
357
            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
358
        elif val_eer is not None:
Anthony Larcher's avatar
Anthony Larcher committed
359
360
            self.is_best = val_eer < self.best_eer
            self.best_eer = min(val_eer, self.best_eer)
Anthony Larcher's avatar
Anthony Larcher committed
361
362
363
364
365
            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
366
367


Anthony Larcher's avatar
Anthony Larcher committed
368
class Xtractor(torch.nn.Module):
369
370
371
    """
    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
372

Anthony Larcher's avatar
Anthony Larcher committed
373
374
375
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
376
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
377
                 norm_embedding=False,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
378
379
                 aam_margin=0.2,
                 aam_s=30):
Anthony Larcher's avatar
Anthony Larcher committed
380
381
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
382
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
383
        """
Anthony Larcher's avatar
Anthony Larcher committed
384
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
385
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
386
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
387
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
388

Anthony Larcher's avatar
Anthony Larcher committed
389
390
        print(f"Speaker number : {self.speaker_number}")

Anthony Larcher's avatar
Anthony Larcher committed
391
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
392

Anthony Larcher's avatar
Anthony Larcher committed
393
394
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
395
396
397
398
399
            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
400
401
            self.activation = torch.nn.LeakyReLU(0.2)

Anthony Larcher's avatar
Anthony Larcher committed
402
            self.preprocessor = MfccFrontEnd()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
403
            self.feature_size = self.preprocessor.n_mfcc
Anthony Larcher's avatar
Anthony Larcher committed
404

Anthony Larcher's avatar
xv    
Anthony Larcher committed
405
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
406
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
407
                ("activation1", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
408
                ("batch_norm1", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
409
410
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
411
                ("batch_norm2", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
412
413
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
414
                ("batch_norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
415
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
416
                ("activation4", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
417
                ("batch_norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
418
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
419
                ("activation5", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
420
                ("batch_norm5", torch.nn.BatchNorm1d(1536))
Anthony Larcher's avatar
Anthony Larcher committed
421
422
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
423
424
            self.embedding_size = 512

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
425
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
426
            self.stat_pooling_weight_decay = 0
Anthony Larcher's avatar
xv    
Anthony Larcher committed
427
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
428
                ("linear6", torch.nn.Linear(3072, self.embedding_size))
Anthony Larcher's avatar
Anthony Larcher committed
429
430
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
431
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
432
                self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
Anthony Larcher's avatar
Anthony Larcher committed
433
434
435
                                                                int(self.speaker_number),
                                                                s=64,
                                                                m=0.2,
Anthony Larcher's avatar
Anthony Larcher committed
436
                                                                easy_margin=False)
Anthony Larcher's avatar
Anthony Larcher committed
437
438
439
            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
440
                    ("batch_norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
441
442
443
                    ("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
444
                    ("batch_norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
445
446
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
447

448
            self.preprocessor_weight_decay = 0.0002
Anthony Larcher's avatar
debug    
Anthony Larcher committed
449
450
451
            self.sequence_network_weight_decay = 0.0002
            self.before_speaker_embedding_weight_decay = 0.002
            self.after_speaker_embedding_weight_decay = 0.002
Anthony Larcher's avatar
Anthony Larcher committed
452

Anthony Larcher's avatar
Anthony Larcher committed
453
        elif model_archi == "resnet34":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
454

Anthony Larcher's avatar
Anthony Larcher committed
455
            self.preprocessor = MelSpecFrontEnd(n_mels=80)
Anthony Larcher's avatar
Anthony Larcher committed
456
            self.sequence_network = PreResNet34()
Anthony Larcher's avatar
Anthony Larcher committed
457
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
458

Anthony Larcher's avatar
Anthony Larcher committed
459
460
            self.before_speaker_embedding = torch.nn.Linear(in_features=5120,
                                                            out_features=self.embedding_size)
Anthony Larcher's avatar
Anthony Larcher committed
461

Anthony Larcher's avatar
Anthony Larcher committed
462
            self.stat_pooling = AttentivePooling(256, 80, global_context=True)
Anthony Larcher's avatar
Anthony Larcher committed
463

464
            self.loss = "aam"
Anthony Larcher's avatar
Anthony Larcher committed
465
            self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
Anthony Larcher's avatar
Anthony Larcher committed
466
467
                                                            int(self.speaker_number),
                                                            s = 30.0,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
468
                                                            m = 0.20,
Anthony Larcher's avatar
Anthony Larcher committed
469
                                                            easy_margin = False)
Anthony Larcher's avatar
Anthony Larcher committed
470

Anthony Larcher's avatar
Anthony Larcher committed
471
472
473
474
475
            self.preprocessor_weight_decay = 0.00002
            self.sequence_network_weight_decay = 0.00002
            self.stat_pooling_weight_decay = 0.00002
            self.before_speaker_embedding_weight_decay = 0.00002
            self.after_speaker_embedding_weight_decay = 0.0002
Anthony Larcher's avatar
Anthony Larcher committed
476

Anthony Larcher's avatar
Anthony Larcher committed
477
        elif model_archi == "fastresnet34":
Anthony Larcher's avatar
Anthony Larcher committed
478
            self.preprocessor = MelSpecFrontEnd()
Anthony Larcher's avatar
Anthony Larcher committed
479
            self.sequence_network = PreFastResNet34()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
480
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
481

Anthony Larcher's avatar
Anthony Larcher committed
482
            self.before_speaker_embedding = torch.nn.Linear(in_features = 2560,
Anthony Larcher's avatar
Anthony Larcher committed
483
                                                            out_features = self.embedding_size)
Anthony Larcher's avatar
Anthony Larcher committed
484

Anthony Larcher's avatar
Anthony Larcher committed
485
            self.stat_pooling = AttentivePooling(128, 80, global_context=False)
Anthony Larcher's avatar
Anthony Larcher committed
486
487
            self.stat_pooling_weight_decay = 0

Anthony Larcher's avatar
merge    
Anthony Larcher committed
488
489
490
491
            self.loss = loss
            if self.loss == "aam":
                self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
                                                                int(self.speaker_number),
Anthony Larcher's avatar
merge    
Anthony Larcher committed
492
493
                                                                s = 30,
                                                                m = 0.2,
Anthony Larcher's avatar
merge    
Anthony Larcher committed
494
495
496
497
                                                                easy_margin = False)

            elif self.loss == 'aps':
                self.after_speaker_embedding = SoftmaxAngularProto(int(self.speaker_number))
Anthony Larcher's avatar
Anthony Larcher committed
498
499
            elif self.loss == 'smn':
                self.after_speaker_embedding = AngularProximityMagnet(int(self.speaker_number))
Anthony Larcher's avatar
Anthony Larcher committed
500

Anthony Larcher's avatar
Anthony Larcher committed
501
502
503
504
505
            self.preprocessor_weight_decay = 0.00002
            self.sequence_network_weight_decay = 0.00002
            self.stat_pooling_weight_decay = 0.00002
            self.before_speaker_embedding_weight_decay = 0.00002
            self.after_speaker_embedding_weight_decay = 0.0002
Anthony Larcher's avatar
Anthony Larcher committed
506

Anthony Larcher's avatar
Anthony Larcher committed
507
508
        elif model_archi == "halfresnet34":
            self.preprocessor = MelSpecFrontEnd(n_fft=1024,
Anthony Larcher's avatar
Anthony Larcher committed
509
510
                                                win_length=400,
                                                hop_length=160,
Anthony Larcher's avatar
Anthony Larcher committed
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
                                                n_mels=80)
            self.sequence_network = PreHalfResNet34()
            self.embedding_size = 256
            self.before_speaker_embedding = torch.nn.Linear(in_features = 5120,
                                                            out_features = self.embedding_size)
            self.stat_pooling = AttentivePooling(256, 80, global_context=True)

            self.loss = loss
            if self.loss == "aam":
                self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
                                                                int(self.speaker_number),
                                                                s = 30,
                                                                m = 0.2,
                                                                easy_margin = False)
            elif self.loss == 'aps':
                self.after_speaker_embedding = SoftmaxAngularProto(int(self.speaker_number))
            self.preprocessor_weight_decay = 0.00002
            self.sequence_network_weight_decay = 0.00002
            self.stat_pooling_weight_decay = 0.00002
            self.before_speaker_embedding_weight_decay = 0.00002
            self.after_speaker_embedding_weight_decay = 0.0002

Anthony Larcher's avatar
Anthony Larcher committed
533
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
534
535
536
537
538
539

            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
540
541
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
542
543
544
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
545
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
546
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
547
548
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565

            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
566
567
568
569
570
            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
571
            elif self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
572
573
574
                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
575

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
576
577
578
579
580
581
            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
582
        else:
Anthony Larcher's avatar
Anthony Larcher committed
583
584
            is_first_resblock = True

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
585
586
587
588
589
590
            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
591

Anthony Larcher's avatar
Anthony Larcher committed
592
            self.loss = cfg["loss"]["type"]
Anthony Larcher's avatar
Anthony Larcher committed
593
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
594
595
                self.aam_margin = cfg["loss"]["aam_margin"]
                self.aam_s = cfg["loss"]["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
596

Anthony Larcher's avatar
Anthony Larcher committed
597
598
599
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
600
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
601
602
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
603
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
604
605
606
607
608
609
610
611
612
613
614
615
                        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
616
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
617
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
618
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
619
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
620
621
622
623
624
625
                                                        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
626
                self.preprocessor_weight_decay = 0.000
Anthony Larcher's avatar
Anthony Larcher committed
627
628

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
629
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
630
            """
Anthony Larcher's avatar
Anthony Larcher committed
631
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
632
            if self.feature_size is None:
Anthony Larcher's avatar
Anthony Larcher committed
633
                self.feature_size = cfg["preprocessor"]["feature_size"]
Anthony Larcher's avatar
Anthony Larcher committed
634

Anthony Larcher's avatar
Anthony Larcher committed
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
            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
650
651
652
653
654
                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
655
656
657
658
659
                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
660
                                                                stride=1)))
Anthony Larcher's avatar
Anthony Larcher committed
661

Anthony Larcher's avatar
Anthony Larcher committed
662
                elif k.startswith("conv"):
Anthony Larcher's avatar
Anthony Larcher committed
663
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
664
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
665
666
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
667
668
                    input_size = cfg["segmental"][k]["output_channels"]

Anthony Larcher's avatar
Anthony Larcher committed
669
670
                elif k.startswith("ctrans"):
                    segmental_layers.append((k, torch.nn.ConvTranspose1d(input_size,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
671
                                                                         cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
672
673
                                                                         kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                         dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
674
675
676
                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

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

Anthony Larcher's avatar
Anthony Larcher committed
680
681
682
683
684
685
                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
686
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
687
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
688

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
689
690
691
692
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
693
            tmp_input_size = input_size * 2
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
694
695
696
697
            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
698
                tmp_input_size = cfg["stat_pooling"]["gru_node"]
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
699

Anthony Larcher's avatar
Anthony Larcher committed
700
701
            self.stat_pooling_weight_decay = cfg["stat_pooling"]["weight_decay"]

Anthony Larcher's avatar
Anthony Larcher committed
702
            """
Anthony Larcher's avatar
Anthony Larcher committed
703
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
704
            """
Anthony Larcher's avatar
Anthony Larcher committed
705
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
706
            input_size = tmp_input_size
Anthony Larcher's avatar
xv    
Anthony Larcher committed
707
708
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
709
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
710
711
                    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
712
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
713
714
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
715
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
716
717

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
718
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
719

Anthony Larcher's avatar
Anthony Larcher committed
720
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
721
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
722
723

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
724
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
725

Anthony Larcher's avatar
Anthony Larcher committed
726
            self.embedding_size = input_size
Anthony Larcher's avatar
Anthony Larcher committed
727
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
728
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
729

Anthony Larcher's avatar
Anthony Larcher committed
730
            # if loss_criteria is "cce"
Anthony Larcher's avatar
xv    
Anthony Larcher committed
731
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
732
733
734
735
736
737
738
739
            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
740
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
741
                            input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
742

Anthony Larcher's avatar
Anthony Larcher committed
743
744
745
746
747
                    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
748

Anthony Larcher's avatar
Anthony Larcher committed
749
750
                    elif k.startswith("activation"):
                        after_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
751

Anthony Larcher's avatar
Anthony Larcher committed
752
753
                    elif k.startswith('batch_norm'):
                        after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
754

Anthony Larcher's avatar
Anthony Larcher committed
755
756
757
758
759
760
                    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
761
762
763
764
765
766
767
                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
Anthony Larcher committed
768
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
769

Anthony Larcher's avatar
Anthony Larcher committed
770
    def forward(self, x, is_eval=False, target=None, norm_embedding=True):
771
772
773
        """

        :param x:
774
        :param is_eval: False for training
775
776
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
777
        if self.preprocessor is not None:
Anthony Larcher's avatar
merge    
Anthony Larcher committed
778
            x = self.preprocessor(x, is_eval)
Anthony Larcher's avatar
Anthony Larcher committed
779

Anthony Larcher's avatar
Anthony Larcher committed
780
        x = self.sequence_network(x)
781

Anthony Larcher's avatar
Anthony Larcher committed
782
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
783
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
784
785

        x = self.before_speaker_embedding(x)
786

Anthony Larcher's avatar
Anthony Larcher committed
787
        if norm_embedding:
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
788
            x = l2_norm(x)
Anthony Larcher's avatar
Anthony Larcher committed
789

Anthony Larcher's avatar
Anthony Larcher committed
790
        if self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
791
            if is_eval:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
792
                return x
Anthony Larcher's avatar
debug    
Anthony Larcher committed
793
            else:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
794
                return self.after_speaker_embedding(x), x
Anthony Larcher's avatar
Anthony Larcher committed
795

Anthony Larcher's avatar
merge    
Anthony Larcher committed
796
        elif self.loss in ['aam', 'aps']:
Anthony Larcher's avatar
merge    
Anthony Larcher committed
797
798
799
800
801
            x = self.after_speaker_embedding(x, target=target), torch.nn.functional.normalize(x, dim=1)
        elif self.loss == 'smn':
            if not is_eval:
                x = self.after_speaker_embedding(x, target=target), x

Anthony Larcher's avatar
Anthony Larcher committed
802

Anthony Larcher's avatar
Anthony Larcher committed
803
        return x
Anthony Larcher's avatar
Anthony Larcher committed
804

805
806
    def context_size(self):
        context = 1
Anthony Larcher's avatar
Anthony Larcher committed
807
808
809
810
811
812
813
814
815
        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
816

Anthony Larcher's avatar
Anthony Larcher committed
817

Anthony Larcher's avatar
Anthony Larcher committed
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
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):
                fill_dict(v1, source_dict[k1], prefix + "\t")
            else:
                pass
        else:
            if k1 in source_dict and source_dict[k1] is not None:
                target_dict[k1] = source_dict[k1]
            else:
                pass


Anthony Larcher's avatar
Anthony Larcher committed
840
841
def update_training_dictionary(dataset_description,
                               model_description,
Anthony Larcher's avatar
Anthony Larcher committed
842
843
                               training_description,
                               kwargs=None):
Anthony Larcher's avatar
Anthony Larcher committed
844
845
    """

Anthony Larcher's avatar
Anthony Larcher committed
846
847
848
849
    :param dataset_description:
    :param model_description:
    :param training_description:
    :param kwargs:
Anthony Larcher's avatar
Anthony Larcher committed
850
851
852
853
854
855
    :return:
    """
    dataset_opts=dict()
    model_opts=dict()
    training_opts=dict()

Anthony Larcher's avatar
Anthony Larcher committed
856
857
858
859
860
    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
861

Anthony Larcher's avatar
Anthony Larcher committed
862
863
864
865
866
    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
867

Anthony Larcher's avatar
Anthony Larcher committed
868
869
870
871
872
873
874
875
876
    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
Anthony Larcher's avatar
Anthony Larcher committed
877
    dataset_opts["stratify"] = False
Anthony Larcher's avatar
Anthony Larcher committed
878
879
880
881
882
883
884
    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()
Anthony Larcher's avatar
Anthony Larcher committed
885
    dataset_opts["train"]["duration"] = 4.
Anthony Larcher's avatar
Anthony Larcher committed
886
    dataset_opts["train"]["chunk_per_segment"] = -1
Anthony Larcher's avatar
Anthony Larcher committed
887
    dataset_opts["train"]["overlap"] = 3.9
Anthony Larcher's avatar
Anthony Larcher committed
888
889
890
    dataset_opts["train"]["sampler"] = dict()
    dataset_opts["train"]["sampler"]["examples_per_speaker"] = 1
    dataset_opts["train"]["sampler"]["samples_per_speaker"] = 100
Anthony Larcher's avatar
debug    
Anthony Larcher committed
891
    dataset_opts["train"]["sampler"]["augmentation_replica"] = 1
Anthony Larcher's avatar
Anthony Larcher committed
892
    dataset_opts["train"]["transform_number"] = 2
Anthony Larcher's avatar
Anthony Larcher committed
893
894
895
896
897
898
    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()
Anthony Larcher's avatar
Anthony Larcher committed
899
    dataset_opts["train"]["transformation"]["add_reverb"]["rir_db_csv"] = ""
Anthony Larcher's avatar
Anthony Larcher committed
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
    dataset_opts["train"]["transformation"]["add_reverb"]["data_path"] = ""

    dataset_opts["valid"] = dict()
    dataset_opts["valid"]["duration"] = 2.
    dataset_opts["valid"]["transformation"] = dict()
    dataset_opts["valid"]["transformation"]["pipeline"] = ""
    dataset_opts["valid"]["transformation"]["add_noise"] = dict()
    dataset_opts["valid"]["transformation"]["add_noise"]["noise_db_csv"] = ""
    dataset_opts["valid"]["transformation"]["add_noise"]["data_path"] = ""
    dataset_opts["valid"]["transformation"]["add_reverb"] = dict()
    dataset_opts["valid"]["transformation"]["add_reverb"]["noise_db_csv"] = ""
    dataset_opts["valid"]["transformation"]["add_reverb"]["data_path"] = ""

    dataset_opts["test"] = dict()
    dataset_opts["test"]["idmap"] = ""
    dataset_opts["test"]["ndx"] = ""
    dataset_opts["test"]["key"] = ""
    dataset_opts["test"]["data_path"] =""

    # Initialize model options
    model_opts["speaker_number"] = None
    model_opts["loss"] = dict()
    model_opts["loss"]["type"] ="aam"
    model_opts["loss"]["aam_margin"] = 0.2
    model_opts["loss"]["aam_s"] = 30

    model_opts["initial_model_name"] = None
    model_opts["reset_parts"] = []
    model_opts["freeze_parts"] = []

    model_opts["model_type"] = "fastresnet"

Anthony Larcher's avatar
Anthony Larcher committed
932
933
934
    model_opts["preprocessor"] = dict()
    model_opts["preprocessor"]["type"] =  "mel_spec"
    model_opts["preprocessor"]["feature_size"] = 80
Anthony Larcher's avatar
Anthony Larcher committed
935
936
937

    # Initialize training options
    training_opts["log_file"] = "sidekit.log"
Anthony Larcher's avatar
Anthony Larcher committed
938
939
940
    training_opts["numpy_seed"] = 0
    training_opts["torch_seed"] = 0
    training_opts["random_seed"] = 0
Anthony Larcher's avatar
Anthony Larcher committed
941
    training_opts["deterministic"] = False
Anthony Larcher's avatar
Anthony Larcher committed
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
    training_opts["epochs"] = 100
    training_opts["lr"] = 1e-3
    training_opts["patience"] = 50
    training_opts["multi_gpu"] = False
    training_opts["num_cpu"] = 5
    training_opts["mixed_precision"] = False
    training_opts["clipping"] = False

    training_opts["optimizer"] = dict()
    training_opts["optimizer"]["type"] = "sgd"
    training_opts["optimizer"]["options"] = None

    training_opts["scheduler"] = dict()
    training_opts["scheduler"]["type"] = "ReduceLROnPlateau"
    training_opts["scheduler"]["options"] = None

    training_opts["compute_test_eer"] = False
    training_opts["log_interval"] = 10
Anthony Larcher's avatar
Anthony Larcher committed
960
    training_opts["validation_frequency"] = 1
Anthony Larcher's avatar
Anthony Larcher committed
961
962
963
964
965
966
967
968
969
970
971

    training_opts["tmp_model_name"] = "tmp_model.pt"
    training_opts["best_model_name"] = "best_model.pt"
    training_opts["checkpoint_frequency"] = "10"

    # Use options from the YAML config files
    fill_dict(dataset_opts, tmp_data_dict)
    fill_dict(model_opts, tmp_model_dict)
    fill_dict(training_opts, tmp_train_dict)

    # Overwrite with manually given parameters
Anthony Larcher's avatar
Anthony Larcher committed
972
973
974
975
976
977
978
979
980
981
982
983
    if "lr" in kwargs:
        training_opts["lr"] = kwargs['lr']
    if "batch_size" in kwargs:
        dataset_opts["batch_size"] = kwargs["batch_size"]
    if "optimizer" in kwargs:
        training_opts["optimizer"]["type"] = kwargs["optimizer"]
    if "scheduler" in kwargs:
        training_opts["scheduler"]["type"] = kwargs["scheduler"]
    if "margin" in kwargs:
        model_opts["loss"]["aam_margin"] = kwargs["margin"]
    if "aam_s" in kwargs:
        model_opts["loss"]["aam_s"] = kwargs["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
984
985
986
987

    return dataset_opts, model_opts, training_opts


Anthony Larcher's avatar
debug    
Anthony Larcher committed
988
def get_network(model_opts, local_rank):
Anthony Larcher's avatar
Anthony Larcher committed
989
    """
Anthony Larcher's avatar
Anthony Larcher committed
990
991

    :param model_opts:
Anthony Larcher's avatar
Anthony Larcher committed
992
    :param local_rank:
Anthony Larcher's avatar
Anthony Larcher committed
993
    :return:
Anthony Larcher's avatar
Anthony Larcher committed
994
995
    """

Anthony Larcher's avatar
Anthony Larcher committed
996
    if model_opts["model_type"] in ["xvector", "rawnet2", "resnet34", "fastresnet34", "halfresnet34"]:
Anthony Larcher's avatar
Anthony Larcher committed
997
        model = Xtractor(model_opts["speaker_number"], model_opts["model_type"], loss=model_opts["loss"]["type"])
Anthony Larcher's avatar
Anthony Larcher committed
998
    else:
Anthony Larcher's avatar
Anthony Larcher committed
999
        # Custom type of model
Anthony Larcher's avatar
Anthony Larcher committed
1000
        model = Xtractor(model_opts["speaker_number"], model_opts, loss=model_opts["loss"]["type"])
Anthony Larcher's avatar
Anthony Larcher committed
1001

Anthony Larcher's avatar
Anthony Larcher committed
1002
1003
1004
1005
    # Load the model if it exists
    if model_opts["initial_model_name"] is not None and os.path.isfile(model_opts["initial_model_name"]):
        logging.critical(f"*** Load model from = {model_opts['initial_model_name']}")
        checkpoint = torch.load(model_opts["initial_model_name"])
Anthony Larcher's avatar
Anthony Larcher committed
1006

Anthony Larcher's avatar
Anthony Larcher committed
1007
1008
        """
        Here we remove all layers that we don't want to reload
Anthony Larcher's avatar
Anthony Larcher committed
1009

Anthony Larcher's avatar
Anthony Larcher committed
1010
1011
1012
1013
        """
        pretrained_dict = checkpoint["model_state_dict"]
        for part in model_opts["reset_parts"]:
            pretrained_dict = {k: v for k, v in pretrained_dict.items() if not k.startswith(part)}
Anthony Larcher's avatar
Anthony Larcher committed
1014

Anthony Larcher's avatar
Anthony Larcher committed
1015
1016
1017
        new_model_dict = model.state_dict()
        new_model_dict.update(pretrained_dict)
        model.load_state_dict(new_model_dict)
Anthony Larcher's avatar
Anthony Larcher committed
1018
1019
1020

        # Freeze required layers
        for name, param in model.named_parameters():
Anthony Larcher's avatar
Anthony Larcher committed
1021
            if name.split(".")[0] in model_opts["reset_parts"]:
Anthony Larcher's avatar
Anthony Larcher committed
1022
1023
                param.requires_grad = False

Anthony Larcher's avatar
clean    
Anthony Larcher committed
1024
1025
1026
1027
    if model_opts["loss"]["type"] == "aam" and not (model_opts["loss"]["aam_margin"] == 0.2 and model_opts["loss"]["aam_s"] == 30):
        model.after_speaker_embedding.change_params(model_opts["loss"]["aam_s"], model_opts["loss"]["aam_margin"])
        print(f"Modified AAM: margin = {model.after_speaker_embedding.m} and s = {model.after_speaker_embedding.s}")

Anthony Larcher's avatar
debug    
Anthony Larcher committed
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
    if local_rank < 1:
        logging.info(model)
        logging.info("Model_parameters_count: {:d}".format(
            sum(p.numel()
                for p in model.sequence_network.parameters()
                if p.requires_grad) + \
            sum(p.numel()
                for p in model.before_speaker_embedding.parameters()
                if p.requires_grad) + \
            sum(p.numel()
                for p in model.stat_pooling.parameters()
                if p.requires_grad)))
Anthony Larcher's avatar
Anthony Larcher committed
1040
1041
1042
1043

    return model


Anthony Larcher's avatar
Anthony Larcher committed
1044
def get_loaders(dataset_opts, training_opts, model_opts, local_rank=0):
Anthony Larcher's avatar
Anthony Larcher committed
1045
1046
    """

Anthony Larcher's avatar
Anthony Larcher committed
1047
1048
1049
    :param dataset_opts:
    :param training_opts:
    :param model_opts:
Anthony Larcher's avatar
Anthony Larcher committed
1050
1051
1052
1053
1054
1055
1056
    :return:
    """

    """
    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
Anthony Larcher's avatar
Anthony Larcher committed
1057
    Then we provide those two
Anthony Larcher's avatar
Anthony Larcher committed
1058
    """
Anthony Larcher's avatar
Anthony Larcher committed
1059
    df = pandas.read_csv(dataset_opts["dataset_csv"])
Anthony Larcher's avatar
Anthony Larcher committed
1060

Anthony Larcher's avatar
Anthony Larcher committed
1061
1062
1063
    stratify = None
    if dataset_opts["stratify"]:
        stratify = df["speaker_idx"]
Anthony Larcher's avatar
Anthony Larcher committed
1064
1065