xvector.py 89.3 KB
Newer Older
Anthony Larcher's avatar
Anthony Larcher committed
1
# coding: utf-8 -*-
Anthony Larcher's avatar
Anthony Larcher committed
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
#
# This file is part of SIDEKIT.
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is free software: you can redistribute it and/or modify
# it under the terms of the GNU LLesser General Public License as
# published by the Free Software Foundation, either version 3 of the License,
# or (at your option) any later version.
#
# SIDEKIT is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU Lesser General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with SIDEKIT.  If not, see <http://www.gnu.org/licenses/>.

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

Anthony Larcher's avatar
Anthony Larcher committed
28

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

Anthony Larcher's avatar
Anthony Larcher committed
43
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
44
45
from torch.utils.data import DataLoader
from sklearn.model_selection import train_test_split
Anthony Larcher's avatar
Anthony Larcher committed
46
47
48
from .pooling import MeanStdPooling
from .pooling import AttentivePooling
from .pooling import GruPooling
Anthony Larcher's avatar
Anthony Larcher committed
49
50
51
from .preprocessor import MfccFrontEnd
from .preprocessor import MelSpecFrontEnd
from .preprocessor import RawPreprocessor
Anthony Larcher's avatar
Anthony Larcher committed
52
from .xsets import SideSet
Anthony Larcher's avatar
Anthony Larcher committed
53
from .xsets import IdMapSet
Anthony Larcher's avatar
Anthony Larcher committed
54
from .xsets import IdMapSetPerSpeaker
Anthony Larcher's avatar
Anthony Larcher committed
55
from .xsets import SideSampler
Anthony Larcher's avatar
Anthony Larcher committed
56
57
58
59
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
60
from ..bosaris import IdMap
Anthony Larcher's avatar
Anthony Larcher committed
61
62
from ..bosaris import Key
from ..bosaris import Ndx
Anthony Larcher's avatar
Anthony Larcher committed
63
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
64
from ..iv_scoring import cosine_scoring
Anthony Larcher's avatar
Anthony Larcher committed
65
from .sincnet import SincNet
Anthony Larcher's avatar
merge    
Anthony Larcher committed
66
from ..bosaris.detplot import rocch, rocch2eer
Anthony Larcher's avatar
merge    
Anthony Larcher committed
67
from .loss import SoftmaxAngularProto, ArcLinear
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
68
from .loss import l2_norm
Anthony Larcher's avatar
Anthony Larcher committed
69
from .loss import ArcMarginProduct
Anthony Larcher's avatar
Anthony Larcher committed
70

Anthony Larcher's avatar
Anthony Larcher committed
71
from ..sidekit_io import init_logging
Anthony Larcher's avatar
ddp    
Anthony Larcher committed
72

Anthony Larcher's avatar
Anthony Larcher committed
73
74
torch.backends.cudnn.benchmark = True

Anthony Larcher's avatar
Anthony Larcher committed
75
os.environ['MKL_THREADING_LAYER'] = 'GNU'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
76

Anthony Larcher's avatar
Anthony Larcher committed
77
78
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
v1.3.7    
Anthony Larcher committed
79
__copyright__ = "Copyright 2015-2021 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
80
81
82
83
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
84
85


Anthony Larcher's avatar
Anthony Larcher committed
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
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
200
201
def test_metrics(model,
                 device,
Anthony Larcher's avatar
Anthony Larcher committed
202
                 speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
203
204
                 num_thread,
                 mixed_precision):
Anthony Larcher's avatar
Anthony Larcher committed
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
    """Compute model metrics

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

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

    Returns:
        [type]: [description]
    """
Anthony Larcher's avatar
Anthony Larcher committed
221

Anthony Larcher's avatar
Anthony Larcher committed
222
223
224
225
    idmap_test_filename = '/lium/raid01_c/larcher/data/allies_dev_verif_idmap.h5'
    ndx_test_filename = '/lium/raid01_c/larcher/data/allies_dev_verif_ndx.h5'
    key_test_filename = '/lium/raid01_c/larcher/data/allies_dev_verif_key.h5'
    data_root_name='/lium/corpus/base/ALLIES/wav'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
226

227
    transform_pipeline = dict()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
228
229
230
231
232

    xv_stat = extract_embeddings(idmap_name=idmap_test_filename,
                                 model_filename=model,
                                 data_root_name=data_root_name,
                                 device=device,
Anthony Larcher's avatar
Anthony Larcher committed
233
                                 loss="aam",
Anthony Larcher's avatar
Anthony Larcher committed
234
235
                                 transform_pipeline=transform_pipeline,
                                 num_thread=num_thread,
Anthony Larcher's avatar
Anthony Larcher committed
236
                                 mixed_precision=mixed_precision)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
237

Anthony Larcher's avatar
merge    
Anthony Larcher committed
238
239
240
241
242
243
244
    tar, non = cosine_scoring(xv_stat,
                              xv_stat,
                              Ndx(ndx_test_filename),
                              wccn=None,
                              check_missing=True,
                              device=device
                              ).get_tar_non(Key(key_test_filename))
Anthony Larcher's avatar
debug    
Anthony Larcher committed
245

Anthony Larcher's avatar
merge    
Anthony Larcher committed
246
247
248
249
    #test_eer = eer(numpy.array(non).astype(numpy.double), numpy.array(tar).astype(numpy.double))
    pmiss, pfa = rocch(tar, non)

    return rocch2eer(pmiss, pfa)
Anthony Larcher's avatar
Anthony Larcher committed
250

Anthony Larcher's avatar
Anthony Larcher committed
251
def new_test_metrics(model,
Anthony Larcher's avatar
Anthony Larcher committed
252
                     device,
Anthony Larcher's avatar
Anthony Larcher committed
253
                     model_opts,
Anthony Larcher's avatar
Anthony Larcher committed
254
255
                     data_opts,
                     train_opts):
Anthony Larcher's avatar
Anthony Larcher committed
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
    """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
274
    xv_stat = extract_embeddings(idmap_name=data_opts["test"]["idmap"],
Anthony Larcher's avatar
Anthony Larcher committed
275
                                 model_filename=model,
Anthony Larcher's avatar
Anthony Larcher committed
276
                                 data_root_name=data_opts["test"]["data_path"],
Anthony Larcher's avatar
Anthony Larcher committed
277
                                 device=device,
Anthony Larcher's avatar
Anthony Larcher committed
278
                                 loss=model_opts["loss"]["type"],
Anthony Larcher's avatar
Anthony Larcher committed
279
                                 transform_pipeline=transform_pipeline,
Anthony Larcher's avatar
Anthony Larcher committed
280
                                 num_thread=train_opts["num_cpu"],
Anthony Larcher's avatar
Anthony Larcher committed
281
282
283
284
                                 mixed_precision=train_opts["mixed_precision"])

    tar, non = cosine_scoring(xv_stat,
                              xv_stat,
Anthony Larcher's avatar
Anthony Larcher committed
285
                              Ndx(data_opts["test"]["ndx"]),
Anthony Larcher's avatar
Anthony Larcher committed
286
287
288
                              wccn=None,
                              check_missing=True,
                              device=device
Anthony Larcher's avatar
Anthony Larcher committed
289
                              ).get_tar_non(Key(data_opts["test"]["key"]))
Anthony Larcher's avatar
Anthony Larcher committed
290
291
292
293
294
295

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

    return rocch2eer(pmiss, pfa)

Anthony Larcher's avatar
Anthony Larcher committed
296

Anthony Larcher's avatar
Anthony Larcher committed
297
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
298
299
300
301
302
303
304
305
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
306
307
308
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
309

Anthony Larcher's avatar
Anthony Larcher committed
310

Anthony Larcher's avatar
Anthony Larcher committed
311
class TrainingMonitor():
Anthony Larcher's avatar
Anthony Larcher committed
312
    """
Anthony Larcher's avatar
Anthony Larcher committed
313

Anthony Larcher's avatar
Anthony Larcher committed
314
    """
Anthony Larcher's avatar
Anthony Larcher committed
315
    def __init__(self,
Anthony Larcher's avatar
Anthony Larcher committed
316
                 output_file,
Anthony Larcher's avatar
Anthony Larcher committed
317
                 log_interval=10,
Anthony Larcher's avatar
Anthony Larcher committed
318
319
                 patience=numpy.inf,
                 best_accuracy=0.0,
Anthony Larcher's avatar
Anthony Larcher committed
320
                 best_eer_epoch=1,
Anthony Larcher's avatar
Anthony Larcher committed
321
322
323
                 best_eer=100,
                 compute_test_eer=False
                 ):
Anthony Larcher's avatar
Anthony Larcher committed
324

Anthony Larcher's avatar
Anthony Larcher committed
325
        self.current_epoch = 0
Anthony Larcher's avatar
Anthony Larcher committed
326
        self.log_interval = log_interval
Anthony Larcher's avatar
Anthony Larcher committed
327
328
329
        self.init_patience = patience
        self.current_patience = patience
        self.best_accuracy = best_accuracy
Anthony Larcher's avatar
Anthony Larcher committed
330
        self.best_eer_epoch = best_eer_epoch
Anthony Larcher's avatar
Anthony Larcher committed
331
        self.best_eer = best_eer
Anthony Larcher's avatar
Anthony Larcher committed
332
        self.compute_test_eer = compute_test_eer
Anthony Larcher's avatar
Anthony Larcher committed
333
        self.test_eer = []
Anthony Larcher's avatar
Anthony Larcher committed
334
335
336
337

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

Anthony Larcher's avatar
Anthony Larcher committed
338
339
        self.val_loss = []
        self.val_acc = []
Anthony Larcher's avatar
Anthony Larcher committed
340
341
        self.val_eer = []

Anthony Larcher's avatar
Anthony Larcher committed
342
343
        self.is_best = True

Anthony Larcher's avatar
Anthony Larcher committed
344
        # Initialize the logger
Anthony Larcher's avatar
Anthony Larcher committed
345
        logging_format = '%(asctime)-15s %(message)s'
Anthony Larcher's avatar
Anthony Larcher committed
346
        logging.basicConfig(level=logging.DEBUG, format=logging_format, datefmt='%m-%d %H:%M')
Anthony Larcher's avatar
Anthony Larcher committed
347
348
        self.logger = logging.getLogger('Monitoring')
        self.logger.setLevel(logging.DEBUG)
Anthony Larcher's avatar
Anthony Larcher committed
349
350
        # create file handler which logs even debug messages
        fh = logging.FileHandler(output_file)
Anthony Larcher's avatar
Anthony Larcher committed
351
352
        formatter = logging.Formatter('%(asctime)s - %(levelname)s - %(message)s')
        fh.setFormatter(formatter)
Anthony Larcher's avatar
Anthony Larcher committed
353
        fh.setLevel(logging.DEBUG)
Anthony Larcher's avatar
Anthony Larcher committed
354
        self.logger.addHandler(fh)
Anthony Larcher's avatar
Anthony Larcher committed
355

Anthony Larcher's avatar
Anthony Larcher committed
356
357
358
359
360
361
    def display(self):
        """

        :return:
        """
        # TODO
Anthony Larcher's avatar
Anthony Larcher committed
362
363
        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
364
365
366
367
368
369

    def display_final(self):
        """

        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
370
        self.logger.critical(f"Best accuracy {self.best_accuracy * 100.} obtained at epoch {self.best_accuracy_epoch}")
Anthony Larcher's avatar
Anthony Larcher committed
371
372

    def update(self,
Anthony Larcher's avatar
Anthony Larcher committed
373
               epoch=None,
Anthony Larcher's avatar
Anthony Larcher committed
374
375
               training_acc=None,
               training_loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
376
377
378
379
380
               test_eer=None,
               val_eer=None,
               val_loss=None,
               val_acc=None):

Anthony Larcher's avatar
Anthony Larcher committed
381
382
383
        if epoch is not None:
            self.current_epoch = epoch
        if training_acc is not None:
Anthony Larcher's avatar
Anthony Larcher committed
384
            self.training_acc.append(training_acc)
Anthony Larcher's avatar
Anthony Larcher committed
385
        if training_loss is not None:
Anthony Larcher's avatar
Anthony Larcher committed
386
            self.training_loss.append(training_loss)
Anthony Larcher's avatar
Anthony Larcher committed
387
        if val_eer is not None:
Anthony Larcher's avatar
Anthony Larcher committed
388
            self.val_eer.append(val_eer)
Anthony Larcher's avatar
Anthony Larcher committed
389
        if val_loss is not None:
Anthony Larcher's avatar
Anthony Larcher committed
390
            self.val_loss.append(val_loss)
Anthony Larcher's avatar
Anthony Larcher committed
391
        if val_acc is not None:
Anthony Larcher's avatar
Anthony Larcher committed
392
            self.val_acc.append(val_acc)
Anthony Larcher's avatar
Anthony Larcher committed
393
394

        # remember best accuracy and save checkpoint
Anthony Larcher's avatar
Anthony Larcher committed
395
        if self.compute_test_eer and test_eer is not None:
Anthony Larcher's avatar
Anthony Larcher committed
396
397
398
            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
399
400
401
402
403
            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
404
        elif val_eer is not None:
Anthony Larcher's avatar
Anthony Larcher committed
405
406
            self.is_best = val_eer < self.best_eer
            self.best_eer = min(val_eer, self.best_eer)
Anthony Larcher's avatar
Anthony Larcher committed
407
408
409
410
411
            if self.is_best:
                self.best_eer_epoch = epoch
                self.current_patience = self.init_patience
            else:
                self.current_patience -= 1
Anthony Larcher's avatar
Anthony Larcher committed
412
413


Anthony Larcher's avatar
Anthony Larcher committed
414
class Xtractor(torch.nn.Module):
415
416
417
    """
    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
418

Anthony Larcher's avatar
Anthony Larcher committed
419
420
421
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
422
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
423
                 norm_embedding=False,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
424
425
                 aam_margin=0.2,
                 aam_s=30):
Anthony Larcher's avatar
Anthony Larcher committed
426
427
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
428
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
429
        """
Anthony Larcher's avatar
Anthony Larcher committed
430
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
431
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
432
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
433
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
434

Anthony Larcher's avatar
Anthony Larcher committed
435
436
        print(f"Speaker number : {self.speaker_number}")

Anthony Larcher's avatar
Anthony Larcher committed
437
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
438

Anthony Larcher's avatar
Anthony Larcher committed
439
440
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
441
442
443
444
445
            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
446
447
            self.activation = torch.nn.LeakyReLU(0.2)

Anthony Larcher's avatar
Anthony Larcher committed
448
            self.preprocessor = MfccFrontEnd()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
449
            self.feature_size = self.preprocessor.n_mfcc
Anthony Larcher's avatar
Anthony Larcher committed
450

Anthony Larcher's avatar
xv    
Anthony Larcher committed
451
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
452
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
453
                ("activation1", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
454
                ("batch_norm1", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
455
456
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
457
                ("batch_norm2", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
458
459
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
460
                ("batch_norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
461
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
462
                ("activation4", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
463
                ("batch_norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
464
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
465
                ("activation5", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
466
                ("batch_norm5", torch.nn.BatchNorm1d(1536))
Anthony Larcher's avatar
Anthony Larcher committed
467
468
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
469
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
470
            self.stat_pooling_weight_decay = 0
Anthony Larcher's avatar
xv    
Anthony Larcher committed
471
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
472
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
473
474
            ]))

475
476
            self.embedding_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
477
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
478
479
480
481
482
                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
483
484
485
            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
486
                    ("batch_norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
487
488
489
                    ("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
490
                    ("batch_norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
491
492
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
493

494
            self.preprocessor_weight_decay = 0.0002
Anthony Larcher's avatar
debug    
Anthony Larcher committed
495
496
497
            self.sequence_network_weight_decay = 0.0002
            self.before_speaker_embedding_weight_decay = 0.002
            self.after_speaker_embedding_weight_decay = 0.002
498
            self.embedding_size = 512
Anthony Larcher's avatar
Anthony Larcher committed
499

Anthony Larcher's avatar
Anthony Larcher committed
500
        elif model_archi == "resnet34":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
501

Anthony Larcher's avatar
Anthony Larcher committed
502
            self.preprocessor = MelSpecFrontEnd(n_mels=80)
Anthony Larcher's avatar
Anthony Larcher committed
503
            self.sequence_network = PreResNet34()
Anthony Larcher's avatar
Anthony Larcher committed
504
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
505

Anthony Larcher's avatar
Anthony Larcher committed
506
507
            self.before_speaker_embedding = torch.nn.Linear(in_features=5120,
                                                            out_features=self.embedding_size)
Anthony Larcher's avatar
Anthony Larcher committed
508
509
510
511

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

512
            self.loss = "aam"
Anthony Larcher's avatar
Anthony Larcher committed
513
            self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
Anthony Larcher's avatar
Anthony Larcher committed
514
515
                                                            int(self.speaker_number),
                                                            s = 30.0,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
516
                                                            m = 0.20,
Anthony Larcher's avatar
Anthony Larcher committed
517
                                                            easy_margin = False)
Anthony Larcher's avatar
Anthony Larcher committed
518
519
520
521
522
523
524

            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
525
        elif model_archi == "fastresnet34":
Anthony Larcher's avatar
Anthony Larcher committed
526
            self.preprocessor = MelSpecFrontEnd()
Anthony Larcher's avatar
Anthony Larcher committed
527
            self.sequence_network = PreFastResNet34()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
528
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
529

Anthony Larcher's avatar
Anthony Larcher committed
530
            self.before_speaker_embedding = torch.nn.Linear(in_features = 2560,
Anthony Larcher's avatar
Anthony Larcher committed
531
                                                            out_features = self.embedding_size)
Anthony Larcher's avatar
Anthony Larcher committed
532
533
534
535

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

Anthony Larcher's avatar
merge    
Anthony Larcher committed
536
537
538
539
            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
540
541
                                                                s = 30,
                                                                m = 0.2,
Anthony Larcher's avatar
merge    
Anthony Larcher committed
542
543
544
545
                                                                easy_margin = False)

            elif self.loss == 'aps':
                self.after_speaker_embedding = SoftmaxAngularProto(int(self.speaker_number))
Anthony Larcher's avatar
Anthony Larcher committed
546

Anthony Larcher's avatar
Anthony Larcher committed
547
548
549
550
551
            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
552

Anthony Larcher's avatar
Anthony Larcher committed
553
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
554
555
556
557
558
559

            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
560
561
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
562
563
564
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
565
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
566
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
567
568
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585

            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
586
587
588
589
590
            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
591
            elif self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
592
593
594
                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
595

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
596
597
598
599
600
601
            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
602
        else:
Anthony Larcher's avatar
Anthony Larcher committed
603
604
            is_first_resblock = True

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
605
606
607
608
609
610
            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
611

Anthony Larcher's avatar
Anthony Larcher committed
612
            self.loss = cfg["loss"]["type"]
Anthony Larcher's avatar
Anthony Larcher committed
613
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
614
615
                self.aam_margin = cfg["loss"]["aam_margin"]
                self.aam_s = cfg["loss"]["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
616

Anthony Larcher's avatar
Anthony Larcher committed
617
618
619
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
620
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
621
622
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
623
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
624
625
626
627
628
629
630
631
632
633
634
635
                        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
636
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
637
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
638
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
639
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
640
641
642
643
644
645
                                                        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
646
                self.preprocessor_weight_decay = 0.000
Anthony Larcher's avatar
Anthony Larcher committed
647
648

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
649
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
650
            """
Anthony Larcher's avatar
Anthony Larcher committed
651
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
652
            if self.feature_size is None:
Anthony Larcher's avatar
Anthony Larcher committed
653
                self.feature_size = cfg["preprocessor"]["feature_size"]
Anthony Larcher's avatar
Anthony Larcher committed
654

Anthony Larcher's avatar
Anthony Larcher committed
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
            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
670
671
672
673
674
                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
675
676
677
678
679
                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
680
                                                                stride=1)))
Anthony Larcher's avatar
Anthony Larcher committed
681

Anthony Larcher's avatar
Anthony Larcher committed
682
                elif k.startswith("conv"):
Anthony Larcher's avatar
Anthony Larcher committed
683
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
684
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
685
686
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
687
688
                    input_size = cfg["segmental"][k]["output_channels"]

Anthony Larcher's avatar
Anthony Larcher committed
689
690
                elif k.startswith("ctrans"):
                    segmental_layers.append((k, torch.nn.ConvTranspose1d(input_size,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
691
                                                                         cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
692
693
                                                                         kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                         dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
694
695
696
                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
697
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
Anthony Larcher committed
698
699
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
700
701
702
703
704
705
                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
706
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
707
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
708

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
709
710
711
712
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
713
            tmp_input_size = input_size * 2
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
714
715
716
717
            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
718
                tmp_input_size = cfg["stat_pooling"]["gru_node"]
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
719

Anthony Larcher's avatar
Anthony Larcher committed
720
721
            self.stat_pooling_weight_decay = cfg["stat_pooling"]["weight_decay"]

Anthony Larcher's avatar
Anthony Larcher committed
722
            """
Anthony Larcher's avatar
Anthony Larcher committed
723
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
724
            """
Anthony Larcher's avatar
Anthony Larcher committed
725
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
726
            input_size = tmp_input_size
Anthony Larcher's avatar
xv    
Anthony Larcher committed
727
728
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
729
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
730
731
                    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
732
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
733
734
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
735
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
736
737

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
738
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
739

Anthony Larcher's avatar
Anthony Larcher committed
740
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
741
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
742
743

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
744
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
745

Anthony Larcher's avatar
Anthony Larcher committed
746
            self.embedding_size = input_size
Anthony Larcher's avatar
Anthony Larcher committed
747
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
748
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
749

Anthony Larcher's avatar
Anthony Larcher committed
750
            # if loss_criteria is "cce"
Anthony Larcher's avatar
xv    
Anthony Larcher committed
751
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
752
753
754
755
756
757
758
759
            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
760
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
761
                            input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
762

Anthony Larcher's avatar
Anthony Larcher committed
763
764
765
766
767
                    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
768

Anthony Larcher's avatar
Anthony Larcher committed
769
770
                    elif k.startswith("activation"):
                        after_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
771

Anthony Larcher's avatar
Anthony Larcher committed
772
773
                    elif k.startswith('batch_norm'):
                        after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
774

Anthony Larcher's avatar
Anthony Larcher committed
775
776
777
778
779
780
                    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
781
782
783
784
785
786
787
                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
788
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
789

790
    def forward(self, x, is_eval=False, target=None, extract_after_pooling=False):
791
792
793
        """

        :param x:
794
        :param is_eval: False for training
795
796
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
797
798
799
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
800
        x = self.sequence_network(x)
801

Anthony Larcher's avatar
Anthony Larcher committed
802
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
803
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
804

805
806
807
        if extract_after_pooling:
            return x

Anthony Larcher's avatar
Anthony Larcher committed
808
        x = self.before_speaker_embedding(x)
809

Anthony Larcher's avatar
Anthony Larcher committed
810
        if self.norm_embedding:
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
811
            x = l2_norm(x)
Anthony Larcher's avatar
Anthony Larcher committed
812

Anthony Larcher's avatar
Anthony Larcher committed
813
        if self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
814
            if is_eval:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
815
                return x
Anthony Larcher's avatar
debug    
Anthony Larcher committed
816
            else:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
817
                return self.after_speaker_embedding(x), x
Anthony Larcher's avatar
Anthony Larcher committed
818

Anthony Larcher's avatar
merge    
Anthony Larcher committed
819
        elif self.loss in ['aam', 'aps']:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
820
821
            if is_eval:
                x = torch.nn.functional.normalize(x, dim=1)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
822
            else:
Anthony Larcher's avatar
merge    
Anthony Larcher committed
823
                x = self.after_speaker_embedding(x, target=target), torch.nn.functional.normalize(x, dim=1)
Anthony Larcher's avatar
Anthony Larcher committed
824

Anthony Larcher's avatar
Anthony Larcher committed
825
        return x
Anthony Larcher's avatar
Anthony Larcher committed
826

827
828
    def context_size(self):
        context = 1
Anthony Larcher's avatar
Anthony Larcher committed
829
830
831
832
833
834
835
836
837
        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
838

Anthony Larcher's avatar
Anthony Larcher committed
839

Anthony Larcher's avatar
Anthony Larcher committed
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
def fill_dict(target_dict, source_dict, prefix = ""):
    """
    Recursively Fill a dictionary target_dict by taking values from source_dict

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

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


Anthony Larcher's avatar
Anthony Larcher committed
867
868
def update_training_dictionary(dataset_description,
                               model_description,
Anthony Larcher's avatar
Anthony Larcher committed
869
870
                               training_description,
                               kwargs=None):
Anthony Larcher's avatar
Anthony Larcher committed
871
872
    """

Anthony Larcher's avatar
Anthony Larcher committed
873
874
875
876
    :param dataset_description:
    :param model_description:
    :param training_description:
    :param kwargs:
Anthony Larcher's avatar
Anthony Larcher committed
877
878
879
880
881
882
    :return:
    """
    dataset_opts=dict()
    model_opts=dict()
    training_opts=dict()

Anthony Larcher's avatar
Anthony Larcher committed
883
884
885
886
887
    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
888

Anthony Larcher's avatar
Anthony Larcher committed
889
890
891
892
893
    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
894

Anthony Larcher's avatar
Anthony Larcher committed
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
    if isinstance(training_description, str) and os.path.isfile(training_description):
        with open(training_description, 'r') as fh:
            tmp_train_dict = yaml.load(fh, Loader=yaml.FullLoader)
    else:
        tmp_train_dict = training_description

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

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

    dataset_opts["train"] = dict()
    dataset_opts["train"]["duration"] = 2.
    dataset_opts["train"]["chunk_per_segment"] = -1
    dataset_opts["train"]["overlap"] = 1.9
Anthony Larcher's avatar
Anthony Larcher committed
914
915
916
    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
Anthony Larcher committed
917
918
919
920
921
922
    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
923
    dataset_opts["train"]["transformation"]["add_reverb"]["rir_db_csv"] = ""
Anthony Larcher's avatar
Anthony Larcher committed
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
    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
956
957
958
    model_opts["preprocessor"] = dict()
    model_opts["preprocessor"]["type"] =  "mel_spec"
    model_opts["preprocessor"]["feature_size"] = 80
Anthony Larcher's avatar
Anthony Larcher committed
959
960
961
962

    # Initialize training options
    training_opts["log_file"] = "sidekit.log"
    training_opts["seed"] = 42
Anthony Larcher's avatar
Anthony Larcher committed
963
    training_opts["deterministic"] = False
Anthony Larcher's avatar
Anthony Larcher committed
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
    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
982
    training_opts["validation_frequency"] = 1
Anthony Larcher's avatar
Anthony Larcher committed
983
984
985
986
987
988
989
990
991
992
993
994

    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
995
996
997
998
999
1000
    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"]