xvector.py 78.3 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
Gaël Le Lan's avatar
Gaël Le Lan committed
30
import math
Anthony Larcher's avatar
Anthony Larcher committed
31
import os
Gaël Le Lan's avatar
Gaël Le Lan committed
32
import numpy
Gaël Le Lan's avatar
Gaël Le Lan committed
33
import random
Gaël Le Lan's avatar
Gaël Le Lan 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
import yaml
Anthony Larcher's avatar
Anthony Larcher committed
39
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
40
41
from torch.utils.data import DataLoader
from sklearn.model_selection import train_test_split
Anthony Larcher's avatar
Anthony Larcher committed
42
43
44
from .pooling import MeanStdPooling
from .pooling import AttentivePooling
from .pooling import GruPooling
Anthony Larcher's avatar
Anthony Larcher committed
45
46
47
from .preprocessor import MfccFrontEnd
from .preprocessor import MelSpecFrontEnd
from .preprocessor import RawPreprocessor
Gaël Le Lan's avatar
Gaël Le Lan committed
48
from .xsets import SideSet
Anthony Larcher's avatar
Anthony Larcher committed
49
from .xsets import IdMapSet
Anthony Larcher's avatar
Anthony Larcher committed
50
from .xsets import IdMapSetPerSpeaker
Gaël Le Lan's avatar
Gaël Le Lan committed
51
from .xsets import SideSampler
Anthony Larcher's avatar
Anthony Larcher committed
52
53
54
from .res_net import ResBlockWFMS
from .res_net import ResBlock
from .res_net import PreFastResNet34
Anthony Larcher's avatar
Anthony Larcher committed
55
from .res_net import PreHalfResNet34
Anthony Larcher's avatar
Anthony Larcher committed
56
from .res_net import PreResNet34
Anthony Larcher's avatar
Anthony Larcher committed
57
from ..bosaris import IdMap
Anthony Larcher's avatar
Anthony Larcher committed
58
59
from ..bosaris import Key
from ..bosaris import Ndx
Anthony Larcher's avatar
Anthony Larcher committed
60
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
61
from ..iv_scoring import cosine_scoring
Anthony Larcher's avatar
Anthony Larcher committed
62
from .sincnet import SincNet
Gaël Le Lan's avatar
Gaël Le Lan committed
63
from ..bosaris.detplot import rocch, rocch2eer
Anthony Larcher's avatar
Anthony Larcher committed
64
from .loss import SoftmaxAngularProto
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
65
from .loss import l2_norm
Anthony Larcher's avatar
Anthony Larcher committed
66
from .loss import ArcMarginProduct
Anthony Larcher's avatar
Anthony Larcher committed
67
68
from .loss import ArcLinear
from .loss import AngularProximityMagnet
Anthony Larcher's avatar
Anthony Larcher committed
69

Anthony Larcher's avatar
ddp    
Anthony Larcher committed
70

Anthony Larcher's avatar
Anthony Larcher committed
71
os.environ['MKL_THREADING_LAYER'] = 'GNU'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
72

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

81
torch.backends.cudnn.benchmark = True
Anthony Larcher's avatar
Anthony Larcher committed
82

Anthony Larcher's avatar
debug    
Anthony Larcher committed
83
def seed_worker(seed_val):
Anthony Larcher's avatar
Anthony Larcher committed
84
    """
Anthony Larcher's avatar
Anthony Larcher committed
85

Anthony Larcher's avatar
Anthony Larcher committed
86
87
88
    :param worker_id:
    :return:
    """
Gaël Le Lan's avatar
Gaël Le Lan committed
89
90
91
    worker_seed = torch.initial_seed() % 2**32
    numpy.random.seed(worker_seed)
    random.seed(worker_seed)
Anthony Larcher's avatar
Anthony Larcher committed
92

Anthony Larcher's avatar
Anthony Larcher committed
93

Anthony Larcher's avatar
Anthony Larcher committed
94
def eer(negatives, positives):
Anthony Larcher's avatar
Anthony Larcher committed
95
    """
Anthony Larcher's avatar
Anthony Larcher committed
96
    Logarithmic complexity EER computation
Anthony Larcher's avatar
Anthony Larcher committed
97

Anthony Larcher's avatar
Anthony Larcher committed
98
99
100
101
    :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
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
    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
Anthony Larcher's avatar
Anthony Larcher committed
139

Anthony Larcher's avatar
Anthony Larcher committed
140
    eer_predicate = 100
Anthony Larcher's avatar
Anthony Larcher committed
141

Anthony Larcher's avatar
Anthony Larcher committed
142
143
144
145
    tfr = (abs(p_index))/pos_count
    tfa = (1+abs(n_index))/neg_count
    if (p_score == n_score and tfr == tfa):
        return tfr
Anthony Larcher's avatar
Anthony Larcher committed
146

Anthony Larcher's avatar
Anthony Larcher committed
147
148
149
150
151
152
153
154
155
156
    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
Anthony Larcher's avatar
Anthony Larcher committed
157

Anthony Larcher's avatar
Anthony Larcher committed
158
159
    tfr = (1+p_index)/pos_count
    tfa = (1+n_index)/neg_count
Anthony Larcher's avatar
Anthony Larcher committed
160

Anthony Larcher's avatar
Anthony Larcher committed
161
162
163
164
165
166
    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
Anthony Larcher's avatar
Anthony Larcher committed
167

Anthony Larcher's avatar
Anthony Larcher committed
168
169
170
171
172
    if abs(tfr - tfa) <= eer_predicate:
        eer_predicate = abs(tfr - tfa)
        eer = (tfr + tfa) / 2
    else:
        return eer
Anthony Larcher's avatar
Anthony Larcher committed
173

Anthony Larcher's avatar
Anthony Larcher committed
174
175
176
177
178
    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
Anthony Larcher's avatar
Anthony Larcher committed
179
    else:
Anthony Larcher's avatar
Anthony Larcher committed
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
        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
Anthony Larcher committed
203
204


Gaël Le Lan's avatar
Gaël Le Lan committed
205
206
def test_metrics(model,
                 device,
Anthony Larcher's avatar
Anthony Larcher committed
207
208
209
                 model_opts,
                 data_opts,
                 train_opts):
Gaël Le Lan's avatar
Gaël Le Lan committed
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
    """Compute model metrics

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

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

    Returns:
        [type]: [description]
    """
Gaël Le Lan's avatar
Gaël Le Lan committed
225
226
    transform_pipeline = dict()

Anthony Larcher's avatar
Anthony Larcher committed
227
    xv_stat = extract_embeddings(idmap_name=data_opts["test"]["idmap"],
Gaël Le Lan's avatar
Gaël Le Lan committed
228
                                 model_filename=model,
Anthony Larcher's avatar
Anthony Larcher committed
229
                                 data_root_name=data_opts["test"]["data_path"],
Gaël Le Lan's avatar
Gaël Le Lan committed
230
                                 device=device,
Gaël Le Lan's avatar
Gaël Le Lan committed
231
                                 transform_pipeline=transform_pipeline,
Anthony Larcher's avatar
Anthony Larcher committed
232
                                 num_thread=train_opts["num_cpu"],
Anthony Larcher's avatar
Anthony Larcher committed
233
                                 mixed_precision=train_opts["mixed_precision"])
Gaël Le Lan's avatar
Gaël Le Lan committed
234

Gaël Le Lan's avatar
Gaël Le Lan committed
235
    tar, non = cosine_scoring(xv_stat,
Anthony Larcher's avatar
Anthony Larcher committed
236
                              xv_stat,
Anthony Larcher's avatar
Anthony Larcher committed
237
                              Ndx(data_opts["test"]["ndx"]),
Anthony Larcher's avatar
Anthony Larcher committed
238
239
240
                              wccn=None,
                              check_missing=True,
                              device=device
Anthony Larcher's avatar
Anthony Larcher committed
241
                              ).get_tar_non(Key(data_opts["test"]["key"]))
Anthony Larcher's avatar
Anthony Larcher committed
242

Gaël Le Lan's avatar
Gaël Le Lan committed
243
    pmiss, pfa = rocch(tar, non)
Anthony Larcher's avatar
Anthony Larcher committed
244

Gaël Le Lan's avatar
Gaël Le Lan committed
245
    return rocch2eer(pmiss, pfa)
Gaël Le Lan's avatar
Gaël Le Lan committed
246

Anthony Larcher's avatar
Anthony Larcher committed
247

Anthony Larcher's avatar
Anthony Larcher committed
248
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
249
250
251
252
253
254
255
256
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
257
258
259
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
260

Anthony Larcher's avatar
Anthony Larcher committed
261

Anthony Larcher's avatar
Anthony Larcher committed
262
class TrainingMonitor():
Anthony Larcher's avatar
Anthony Larcher committed
263
264
    """

Gaël Le Lan's avatar
Gaël Le Lan committed
265
    """
Anthony Larcher's avatar
Anthony Larcher committed
266
    def __init__(self,
Anthony Larcher's avatar
Anthony Larcher committed
267
                 output_file,
Anthony Larcher's avatar
Anthony Larcher committed
268
                 log_interval=10,
Anthony Larcher's avatar
Anthony Larcher committed
269
270
                 patience=numpy.inf,
                 best_accuracy=0.0,
Anthony Larcher's avatar
Anthony Larcher committed
271
                 best_eer_epoch=1,
Anthony Larcher's avatar
Anthony Larcher committed
272
273
274
                 best_eer=100,
                 compute_test_eer=False
                 ):
Anthony Larcher's avatar
Anthony Larcher committed
275

Anthony Larcher's avatar
Anthony Larcher committed
276
        self.current_epoch = 0
Anthony Larcher's avatar
Anthony Larcher committed
277
        self.log_interval = log_interval
Anthony Larcher's avatar
Anthony Larcher committed
278
279
280
        self.init_patience = patience
        self.current_patience = patience
        self.best_accuracy = best_accuracy
Anthony Larcher's avatar
Anthony Larcher committed
281
        self.best_eer_epoch = best_eer_epoch
Anthony Larcher's avatar
Anthony Larcher committed
282
        self.best_eer = best_eer
Anthony Larcher's avatar
Anthony Larcher committed
283
        self.compute_test_eer = compute_test_eer
Anthony Larcher's avatar
Anthony Larcher committed
284
        self.test_eer = []
Anthony Larcher's avatar
Anthony Larcher committed
285
286
287
288

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

Anthony Larcher's avatar
Anthony Larcher committed
289
290
        self.val_loss = []
        self.val_acc = []
Anthony Larcher's avatar
Anthony Larcher committed
291
292
        self.val_eer = []

Anthony Larcher's avatar
Anthony Larcher committed
293
294
        self.is_best = True

Anthony Larcher's avatar
Anthony Larcher committed
295
        # Initialize the logger
Anthony Larcher's avatar
Anthony Larcher committed
296
        logging_format = '%(asctime)-15s %(message)s'
Anthony Larcher's avatar
Anthony Larcher committed
297
        logging.basicConfig(level=logging.DEBUG, format=logging_format, datefmt='%m-%d %H:%M')
Anthony Larcher's avatar
Anthony Larcher committed
298
299
        self.logger = logging.getLogger('Monitoring')
        self.logger.setLevel(logging.DEBUG)
Anthony Larcher's avatar
Anthony Larcher committed
300
301
        # create file handler which logs even debug messages
        fh = logging.FileHandler(output_file)
Anthony Larcher's avatar
Anthony Larcher committed
302
303
        formatter = logging.Formatter('%(asctime)s - %(levelname)s - %(message)s')
        fh.setFormatter(formatter)
Anthony Larcher's avatar
Anthony Larcher committed
304
        fh.setLevel(logging.DEBUG)
Anthony Larcher's avatar
Anthony Larcher committed
305
        self.logger.addHandler(fh)
Anthony Larcher's avatar
Anthony Larcher committed
306

Anthony Larcher's avatar
Anthony Larcher committed
307
    def display(self):
Gaël Le Lan's avatar
Gaël Le Lan committed
308
309
310
311
        """

        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
312
        # TODO
Anthony Larcher's avatar
Anthony Larcher committed
313
314
        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
315

Anthony Larcher's avatar
Anthony Larcher committed
316
    def display_final(self):
Anthony Larcher's avatar
Anthony Larcher committed
317
318
319
320
        """

        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
321
        self.logger.critical(f"Best accuracy {self.best_accuracy * 100.} obtained at epoch {self.best_accuracy_epoch}")
Anthony Larcher's avatar
Anthony Larcher committed
322
323

    def update(self,
Anthony Larcher's avatar
Anthony Larcher committed
324
               epoch=None,
Anthony Larcher's avatar
Anthony Larcher committed
325
326
               training_acc=None,
               training_loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
327
328
329
330
331
               test_eer=None,
               val_eer=None,
               val_loss=None,
               val_acc=None):

Anthony Larcher's avatar
Anthony Larcher committed
332
333
334
        if epoch is not None:
            self.current_epoch = epoch
        if training_acc is not None:
Anthony Larcher's avatar
Anthony Larcher committed
335
            self.training_acc.append(training_acc)
Anthony Larcher's avatar
Anthony Larcher committed
336
        if training_loss is not None:
Anthony Larcher's avatar
Anthony Larcher committed
337
            self.training_loss.append(training_loss)
Anthony Larcher's avatar
Anthony Larcher committed
338
        if val_eer is not None:
Anthony Larcher's avatar
Anthony Larcher committed
339
            self.val_eer.append(val_eer)
Anthony Larcher's avatar
Anthony Larcher committed
340
        if val_loss is not None:
Anthony Larcher's avatar
Anthony Larcher committed
341
            self.val_loss.append(val_loss)
Anthony Larcher's avatar
Anthony Larcher committed
342
        if val_acc is not None:
Anthony Larcher's avatar
Anthony Larcher committed
343
            self.val_acc.append(val_acc)
Anthony Larcher's avatar
Anthony Larcher committed
344
345

        # remember best accuracy and save checkpoint
Anthony Larcher's avatar
Anthony Larcher committed
346
        if self.compute_test_eer and test_eer is not None:
Anthony Larcher's avatar
Anthony Larcher committed
347
348
349
            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
350
351
352
353
354
            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
355
        elif val_eer is not None:
Anthony Larcher's avatar
Anthony Larcher committed
356
357
            self.is_best = val_eer < self.best_eer
            self.best_eer = min(val_eer, self.best_eer)
Anthony Larcher's avatar
Anthony Larcher committed
358
359
360
361
362
            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
363

Anthony Larcher's avatar
Anthony Larcher committed
364

Anthony Larcher's avatar
Anthony Larcher committed
365
class Xtractor(torch.nn.Module):
366
367
368
    """
    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
369

Anthony Larcher's avatar
Anthony Larcher committed
370
371
372
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
373
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
374
                 norm_embedding=False,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
375
376
                 aam_margin=0.2,
                 aam_s=30):
Anthony Larcher's avatar
Anthony Larcher committed
377
378
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
379
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
380
        """
Anthony Larcher's avatar
Anthony Larcher committed
381
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
382
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
383
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
384
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
385

Anthony Larcher's avatar
Anthony Larcher committed
386
387
        print(f"Speaker number : {self.speaker_number}")

Anthony Larcher's avatar
Anthony Larcher committed
388
        if model_archi == "xvector":
389
390
391
392
393
394
395

            self.input_nbdim = 2

            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
396

Anthony Larcher's avatar
Anthony Larcher committed
397
398
            self.activation = torch.nn.LeakyReLU(0.2)

Gaël Le Lan's avatar
Gaël Le Lan committed
399
400
            self.preprocessor = MfccFrontEnd()
            self.feature_size = self.preprocessor.n_mfcc
Anthony Larcher's avatar
Anthony Larcher committed
401

Anthony Larcher's avatar
xv    
Anthony Larcher committed
402
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
403
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
404
                ("activation1", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
405
                ("batch_norm1", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
406
407
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
408
                ("batch_norm2", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
409
410
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
411
                ("batch_norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
412
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
413
                ("activation4", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
414
                ("batch_norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
415
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
416
                ("activation5", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
417
                ("batch_norm5", torch.nn.BatchNorm1d(1536))
Anthony Larcher's avatar
Anthony Larcher committed
418
419
            ]))

Gaël Le Lan's avatar
Gaël Le Lan committed
420
421
            self.embedding_size = 512

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
422
            self.stat_pooling = MeanStdPooling()
Gaël Le Lan's avatar
Gaël Le Lan committed
423
            self.stat_pooling_weight_decay = 0
Anthony Larcher's avatar
xv    
Anthony Larcher committed
424
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Gaël Le Lan's avatar
Gaël Le Lan committed
425
                ("linear6", torch.nn.Linear(3072, self.embedding_size))
Anthony Larcher's avatar
Anthony Larcher committed
426
427
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
428
            if self.loss == "aam":
Gaël Le Lan's avatar
Gaël Le Lan committed
429
                self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
Anthony Larcher's avatar
Anthony Larcher committed
430
431
432
                                                                int(self.speaker_number),
                                                                s=64,
                                                                m=0.2,
Gaël Le Lan's avatar
Gaël Le Lan committed
433
                                                                easy_margin=False)
Gaël Le Lan's avatar
Gaël Le Lan committed
434
            elif self.loss == "cce":
Gaël Le Lan's avatar
Gaël Le Lan committed
435
                self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
Gaël Le Lan's avatar
Gaël Le Lan committed
436
                    ("activation6", torch.nn.LeakyReLU(0.2)),
Gaël Le Lan's avatar
Gaël Le Lan committed
437
                    ("batch_norm6", torch.nn.BatchNorm1d(512)),
Gaël Le Lan's avatar
Gaël Le Lan committed
438
                    ("dropout6", torch.nn.Dropout(p=0.05)),
Gaël Le Lan's avatar
Gaël Le Lan committed
439
                    ("linear7", torch.nn.Linear(512, 512)),
Gaël Le Lan's avatar
Gaël Le Lan committed
440
                    ("activation7", torch.nn.LeakyReLU(0.2)),
Gaël Le Lan's avatar
Gaël Le Lan committed
441
442
                    ("batch_norm7", torch.nn.BatchNorm1d(512)),
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
Gaël Le Lan's avatar
Gaël Le Lan committed
443
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
444

Gaël Le Lan's avatar
Gaël Le Lan committed
445
            self.preprocessor_weight_decay = 0.0002
446
447
448
449
450
            self.sequence_network_weight_decay = 0.0002
            self.before_speaker_embedding_weight_decay = 0.002
            self.after_speaker_embedding_weight_decay = 0.002

        elif model_archi == "resnet34":
Gaël Le Lan's avatar
Gaël Le Lan committed
451

Anthony Larcher's avatar
Anthony Larcher committed
452
            self.preprocessor = MelSpecFrontEnd(n_mels=80)
453
            self.sequence_network = PreResNet34()
Anthony Larcher's avatar
Anthony Larcher committed
454
            self.embedding_size = 256
455

Anthony Larcher's avatar
Anthony Larcher committed
456
457
            self.before_speaker_embedding = torch.nn.Linear(in_features=5120,
                                                            out_features=self.embedding_size)
458
459
460
461
462

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

            self.loss = "aam"
Anthony Larcher's avatar
Anthony Larcher committed
463
            self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
464
465
466
                                                            int(self.speaker_number),
                                                            s = 30.0,
                                                            m = 0.20,
Anthony Larcher's avatar
Anthony Larcher committed
467
                                                            easy_margin = False)
468
469

            self.preprocessor_weight_decay = 0.000
Gaël Le Lan's avatar
Gaël Le Lan committed
470
            self.sequence_network_weight_decay = 0.000
471
            self.stat_pooling_weight_decay = 0.000
Gaël Le Lan's avatar
Gaël Le Lan committed
472
473
            self.before_speaker_embedding_weight_decay = 0.00
            self.after_speaker_embedding_weight_decay = 0.00
Anthony Larcher's avatar
Anthony Larcher committed
474

475
        elif model_archi == "fastresnet34":
Gaël Le Lan's avatar
Gaël Le Lan committed
476
            self.preprocessor = MelSpecFrontEnd()
Gaël Le Lan's avatar
Gaël Le Lan committed
477
            self.sequence_network = PreFastResNet34()
Gaël Le Lan's avatar
Gaël Le Lan committed
478
            self.embedding_size = 256
Gaël Le Lan's avatar
Gaël Le Lan committed
479

Gaël Le Lan's avatar
Gaël Le Lan committed
480
            self.before_speaker_embedding = torch.nn.Linear(in_features = 2560,
Gaël Le Lan's avatar
Gaël Le Lan committed
481
                                                            out_features = self.embedding_size)
Anthony Larcher's avatar
Anthony Larcher committed
482

Gaël Le Lan's avatar
Gaël Le Lan committed
483
484
485
486
487
488
489
490
491
492
493
494
            #last_linear = torch.nn.Linear(2560, 1)
            #last_linear.bias.data += 1

            #self.magnitude = torch.nn.Sequential(OrderedDict([
                    #("linear9", torch.nn.Linear(2560, 256)),
                    #("relu9", torch.nn.ReLU()),
                    #("linear10", torch.nn.Linear(256, 256)),
                    #("relu10", torch.nn.ReLU()),
            #        ("linear11", last_linear),
            #        ("relu11", torch.nn.ReLU())
            #        ]))

Gaël Le Lan's avatar
Gaël Le Lan committed
495
            self.stat_pooling = AttentivePooling(128, 80, global_context=False)
Anthony Larcher's avatar
Anthony Larcher committed
496
497
            self.stat_pooling_weight_decay = 0

Gaël Le Lan's avatar
Gaël Le Lan committed
498
499
500
501
            self.loss = loss
            if self.loss == "aam":
                self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
                                                                int(self.speaker_number),
Gaël Le Lan's avatar
Gaël Le Lan committed
502
503
                                                                s = 30.0,
                                                                m = 0.20,
Gaël Le Lan's avatar
Gaël Le Lan committed
504
                                                                easy_margin = False)
Gaël Le Lan's avatar
Gaël Le Lan committed
505

Gaël Le Lan's avatar
Gaël Le Lan committed
506
507
            elif self.loss == 'aps':
                self.after_speaker_embedding = SoftmaxAngularProto(int(self.speaker_number))
Gaël Le Lan's avatar
Gaël Le Lan committed
508
            elif self.loss == 'smn':
Gaël Le Lan's avatar
Gaël Le Lan committed
509
                self.after_speaker_embedding = AngularProximityMagnet(int(self.speaker_number))
Gaël Le Lan's avatar
Gaël Le Lan committed
510

Gaël Le Lan's avatar
Gaël Le Lan committed
511
512
513
514
515
            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
516

Gaël Le Lan's avatar
Gaël Le Lan committed
517
        elif model_archi == "halfresnet34":
Anthony Larcher's avatar
Anthony Larcher committed
518
            self.preprocessor = MelSpecFrontEnd(n_fft=1024,
Gaël Le Lan's avatar
Gaël Le Lan committed
519
520
                                                win_length=400,
                                                hop_length=160,
Gaël Le Lan's avatar
bugfix    
Gaël Le Lan committed
521
                                                n_mels=80)
Gaël Le Lan's avatar
Gaël Le Lan committed
522
            self.sequence_network = PreHalfResNet34()
Gaël Le Lan's avatar
Gaël Le Lan committed
523
            self.embedding_size = 256
Gaël Le Lan's avatar
Gaël Le Lan committed
524
525
526
527
528
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
                                        ("emb1", torch.nn.Linear(in_features = 5120, out_features = self.embedding_size)),
                                        ("bn1", torch.nn.BatchNorm1d(self.embedding_size))
                                    ]))

Gaël Le Lan's avatar
Gaël Le Lan committed
529
530
            self.stat_pooling = AttentivePooling(256, 80, global_context=True)

Gaël Le Lan's avatar
Gaël Le Lan committed
531
532
533
534
            self.loss = loss
            if self.loss == "aam":
                self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
                                                                int(self.speaker_number),
Gaël Le Lan's avatar
Gaël Le Lan committed
535
536
                                                                s = 30.0,
                                                                m = 0.20,
Gaël Le Lan's avatar
Gaël Le Lan committed
537
538
539
                                                                easy_margin = False)
            elif self.loss == 'aps':
                self.after_speaker_embedding = SoftmaxAngularProto(int(self.speaker_number))
Gaël Le Lan's avatar
Gaël Le Lan committed
540
541
542
543
            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
Gaël Le Lan's avatar
Gaël Le Lan committed
544
            self.after_speaker_embedding_weight_decay = 0.0002
Anthony Larcher's avatar
Anthony Larcher committed
545

Anthony Larcher's avatar
Anthony Larcher committed
546
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
547
548
549
550
551
552

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

553
554
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
555
556
557
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
558
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
559
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
560
561
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578

            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
579
580
581
582
583
            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
584
            elif self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
585
586
587
                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
588

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
589
590
591
592
593
594
            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
595
        else:
Anthony Larcher's avatar
Anthony Larcher committed
596
597
            is_first_resblock = True

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
598
599
600
601
602
603
            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
604

Anthony Larcher's avatar
Anthony Larcher committed
605
            self.loss = cfg["loss"]["type"]
Anthony Larcher's avatar
Anthony Larcher committed
606
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
607
608
                self.aam_margin = cfg["loss"]["aam_margin"]
                self.aam_s = cfg["loss"]["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
609

Anthony Larcher's avatar
Anthony Larcher committed
610
611
612
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
613
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
614
615
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
616
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
617
618
619
620
621
622
623
624
625
626
627
628
                        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
629
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
630
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
631
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
632
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
633
634
635
636
637
638
                                                        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
639
                self.preprocessor_weight_decay = 0.000
Anthony Larcher's avatar
Anthony Larcher committed
640
641

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
642
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
643
            """
Anthony Larcher's avatar
Anthony Larcher committed
644
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
645
            if self.feature_size is None:
Anthony Larcher's avatar
Anthony Larcher committed
646
                self.feature_size = cfg["preprocessor"]["feature_size"]
Anthony Larcher's avatar
Anthony Larcher committed
647

Anthony Larcher's avatar
Anthony Larcher committed
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
            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
663
664
665
666
667
                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
668
669
670
671
672
                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
673
                                                                stride=1)))
Anthony Larcher's avatar
Anthony Larcher committed
674

Anthony Larcher's avatar
Anthony Larcher committed
675
                elif k.startswith("conv"):
Anthony Larcher's avatar
Anthony Larcher committed
676
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
677
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
678
679
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
680
681
                    input_size = cfg["segmental"][k]["output_channels"]

Anthony Larcher's avatar
Anthony Larcher committed
682
683
                elif k.startswith("ctrans"):
                    segmental_layers.append((k, torch.nn.ConvTranspose1d(input_size,
Anthony Larcher's avatar
minor    
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
689
                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
690
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
Anthony Larcher committed
691
692
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
693
694
695
696
697
698
                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
699
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
700
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
701

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
702
703
704
705
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
706
            tmp_input_size = input_size * 2
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
707
708
709
710
            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
711
                tmp_input_size = cfg["stat_pooling"]["gru_node"]
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
712

Anthony Larcher's avatar
Anthony Larcher committed
713
714
            self.stat_pooling_weight_decay = cfg["stat_pooling"]["weight_decay"]

Anthony Larcher's avatar
Anthony Larcher committed
715
            """
Anthony Larcher's avatar
Anthony Larcher committed
716
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
717
            """
Anthony Larcher's avatar
Anthony Larcher committed
718
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
719
            input_size = tmp_input_size
Anthony Larcher's avatar
xv    
Anthony Larcher committed
720
721
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
722
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
723
724
                    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
725
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
726
727
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
728
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
729
730

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
731
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
732

Anthony Larcher's avatar
Anthony Larcher committed
733
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
734
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
735
736

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
737
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
738

Anthony Larcher's avatar
Anthony Larcher committed
739
            self.embedding_size = input_size
Anthony Larcher's avatar
Anthony Larcher committed
740
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
741
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
742

Anthony Larcher's avatar
Anthony Larcher committed
743
            # if loss_criteria is "cce"
Anthony Larcher's avatar
xv    
Anthony Larcher committed
744
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
745
746
747
748
749
750
751
752
            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
753
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
754
                            input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
755

Anthony Larcher's avatar
Anthony Larcher committed
756
757
758
759
760
                    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
761

Anthony Larcher's avatar
Anthony Larcher committed
762
763
                    elif k.startswith("activation"):
                        after_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
764

Anthony Larcher's avatar
Anthony Larcher committed
765
766
                    elif k.startswith('batch_norm'):
                        after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
767

Anthony Larcher's avatar
Anthony Larcher committed
768
769
770
771
772
773
                    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
774
                self.norm_embedding = True
Anthony Larcher's avatar
Anthony Larcher committed
775
776
777
778
779
780
                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
781
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
782

783
784
785
786
787
788
789
    def forward(self,
                x,
                is_eval=False,
                target=None,
                extract_after_pooling=False,
                extract_after_preprocessing=False,
                start_after_preprocessing=False):
790
791
792
        """

        :param x:
793
        :param is_eval: False for training
794
795
        :return:
        """
796
        if self.preprocessor is not None and not start_after_preprocessing:
Gaël Le Lan's avatar
Gaël Le Lan committed
797
            x = self.preprocessor(x, is_eval)
Anthony Larcher's avatar
Anthony Larcher committed
798

799
800
801
            if extract_after_preprocessing:
                return x

Anthony Larcher's avatar
Anthony Larcher committed
802
        x = self.sequence_network(x)
803

Anthony Larcher's avatar
Anthony Larcher committed
804
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
805
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
806

807
808
        if extract_after_pooling:
            return x
809

Anthony Larcher's avatar
Anthony Larcher committed
810
        x = self.before_speaker_embedding(x)
811

Anthony Larcher's avatar
Anthony Larcher committed
812
        if self.norm_embedding:
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
813
            x = l2_norm(x)
Gaël Le Lan's avatar
Gaël Le Lan committed
814

Anthony Larcher's avatar
Anthony Larcher committed
815
        if self.loss == "cce":
Gaël Le Lan's avatar
Gaël Le Lan committed
816
            if is_eval:
Gaël Le Lan's avatar
Gaël Le Lan committed
817
                return x
Gaël Le Lan's avatar
Gaël Le Lan committed
818
            else:
Gaël Le Lan's avatar
Gaël Le Lan committed
819
                return self.after_speaker_embedding(x), x
Gaël Le Lan's avatar
Gaël Le Lan committed
820
        elif self.loss in ['aam', 'aps']:
Gaël Le Lan's avatar
Gaël Le Lan committed
821
            x = self.after_speaker_embedding(x, target=target), torch.nn.functional.normalize(x, dim=1)
Gaël Le Lan's avatar
Gaël Le Lan committed
822
823
824
825
        elif self.loss == 'smn':
            if not is_eval:
                x = self.after_speaker_embedding(x, target=target), x

Anthony Larcher's avatar
Anthony Larcher committed
826
        return x
Anthony Larcher's avatar
Anthony Larcher committed
827

828
829
    def context_size(self):
        context = 1
Anthony Larcher's avatar
Anthony Larcher committed
830
831
832
833
834
835
836
837
838
        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
839

Anthony Larcher's avatar
Anthony Larcher committed
840

Anthony Larcher's avatar
Anthony Larcher committed
841
def fill_dict(target_dict, source_dict, prefix = ""):
842
    """
Anthony Larcher's avatar
Anthony Larcher committed
843
    Recursively Fill a dictionary target_dict by taking values from source_dict
844

Anthony Larcher's avatar
Anthony Larcher committed
845
846
    :param target_dict: output dictionary that is initialized with default values
    :param source_dict: input dictionary
847
848
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
849
    for k1, v1 in target_dict.items():
Anthony Larcher's avatar
Anthony Larcher committed
850

Anthony Larcher's avatar
Anthony Larcher committed
851
852
853
854
855
856
857
858
859
860
        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
861

862

Anthony Larcher's avatar
Anthony Larcher committed
863
864
def update_training_dictionary(dataset_description,
                               model_description,
Anthony Larcher's avatar
Anthony Larcher committed
865
866
                               training_description,
                               kwargs=None):
Anthony Larcher's avatar
Anthony Larcher committed
867
    """
868

Anthony Larcher's avatar
Anthony Larcher committed
869
870
871
872
    :param dataset_description:
    :param model_description:
    :param training_description:
    :param kwargs:
Anthony Larcher's avatar
Anthony Larcher committed
873
874
875
876
877
    :return:
    """
    dataset_opts=dict()
    model_opts=dict()
    training_opts=dict()
878

Anthony Larcher's avatar
Anthony Larcher committed
879
880
881
882
883
    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
884

Anthony Larcher's avatar
Anthony Larcher committed
885
886
887
888
889
    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
890

Anthony Larcher's avatar
Anthony Larcher committed
891
892
893
894
895
896
897
898
899
    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
900
    dataset_opts["stratify"] = False
Anthony Larcher's avatar
Anthony Larcher committed
901
902
903
904
905
906
907
    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
908
    dataset_opts["train"]["duration"] = 4.
Anthony Larcher's avatar
Anthony Larcher committed
909
    dataset_opts["train"]["chunk_per_segment"] = -1
Anthony Larcher's avatar
Anthony Larcher committed
910
    dataset_opts["train"]["overlap"] = 3.9
Anthony Larcher's avatar
Anthony Larcher committed
911
912
913
    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
914
    dataset_opts["train"]["sampler"]["augmentation_replica"] = 1
Anthony Larcher's avatar
Anthony Larcher committed
915
    dataset_opts["train"]["transform_number"] = 2
Anthony Larcher's avatar
Anthony Larcher committed
916
917
918
919
920
921
    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()