xvector.py 62 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 pdb
Anthony Larcher's avatar
minor    
Anthony Larcher committed
33
import pickle
Anthony Larcher's avatar
Anthony Larcher committed
34
import shutil
35
import time
Gaël Le Lan's avatar
Gaël Le Lan committed
36
37
38
39
40
41
import traceback
from collections import OrderedDict

import matplotlib.pyplot as plt
import numpy
import pandas
Anthony Larcher's avatar
Anthony Larcher committed
42
import torch
Anthony Larcher's avatar
debug    
Anthony Larcher committed
43
import torchaudio
Anthony Larcher's avatar
Anthony Larcher committed
44
import tqdm
Anthony Larcher's avatar
Anthony Larcher committed
45
import yaml
Anthony Larcher's avatar
Anthony Larcher committed
46
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
47
48
49
50
51
from torch.utils.data import DataLoader
from sklearn.model_selection import train_test_split
from .preprocessor import MfccFrontEnd
from .preprocessor import MelSpecFrontEnd
from .preprocessor import RawPreprocessor
Gaël Le Lan's avatar
Gaël Le Lan 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
Gaël Le Lan's avatar
Gaël Le Lan committed
55
from .xsets import SideSampler
Anthony Larcher's avatar
Anthony Larcher committed
56
57
58
from .res_net import ResBlockWFMS
from .res_net import ResBlock
from .res_net import PreResNet34
Gaël Le Lan's avatar
Gaël Le Lan committed
59
from .res_net import PreFastResNet34, PreHalfResNet34
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
Gaël Le Lan's avatar
Gaël Le Lan committed
66
from ..bosaris.detplot import rocch, rocch2eer
Gaël Le Lan's avatar
Gaël Le Lan 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
Gaël Le Lan's avatar
Gaël Le Lan committed
70
from torch.cuda.amp import autocast, GradScaler
Gaël Le Lan's avatar
Gaël Le Lan committed
71
import torchaudio
Anthony Larcher's avatar
Anthony Larcher committed
72

Anthony Larcher's avatar
ddp    
Anthony Larcher committed
73

Anthony Larcher's avatar
Anthony Larcher committed
74
os.environ['MKL_THREADING_LAYER'] = 'GNU'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
75

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


Anthony Larcher's avatar
Anthony Larcher committed
85
86
logging.basicConfig(format='%(asctime)s %(message)s')

Anthony Larcher's avatar
Anthony Larcher committed
87
88
89

# Make PyTorch Deterministic
torch.manual_seed(0)
Gaël Le Lan's avatar
Gaël Le Lan committed
90
torch.backends.cudnn.deterministic = False
Gaël Le Lan's avatar
Gaël Le Lan committed
91
torch.backends.cudnn.benchmark = True
Anthony Larcher's avatar
Anthony Larcher committed
92
93
94
95
numpy.random.seed(0)



Anthony Larcher's avatar
Anthony Larcher committed
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
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
Anthony Larcher's avatar
Anthony Larcher committed
144

Anthony Larcher's avatar
Anthony Larcher committed
145
    eer_predicate = 100
Anthony Larcher's avatar
Anthony Larcher committed
146

Anthony Larcher's avatar
Anthony Larcher committed
147
148
149
150
    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
151

Anthony Larcher's avatar
Anthony Larcher committed
152
153
154
155
156
157
158
159
160
161
    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
162

Anthony Larcher's avatar
Anthony Larcher committed
163
164
    tfr = (1+p_index)/pos_count
    tfa = (1+n_index)/neg_count
Anthony Larcher's avatar
Anthony Larcher committed
165

Anthony Larcher's avatar
Anthony Larcher committed
166
167
168
169
170
171
    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
172

Anthony Larcher's avatar
Anthony Larcher committed
173
174
175
176
177
    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
178

Anthony Larcher's avatar
Anthony Larcher committed
179
180
181
182
183
    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
184
    else:
Anthony Larcher's avatar
Anthony Larcher committed
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
        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
208
209


Gaël Le Lan's avatar
Gaël Le Lan committed
210
211
def test_metrics(model,
                 device,
Gaël Le Lan's avatar
Gaël Le Lan committed
212
213
214
                 speaker_number,
                 num_thread,
                 mixed_precision):
Gaël Le Lan's avatar
Gaël Le Lan committed
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
    """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]
    """
Gaël Le Lan's avatar
Gaël Le Lan committed
231
232
233
234
235

    idmap_test_filename = 'h5f/vox1_test_cleaned_idmap.h5'
    ndx_test_filename = 'h5f/vox1_test_cleaned_ndx.h5'
    key_test_filename = 'h5f/vox1_test_cleaned_key.h5'
    data_root_name='/hdd/data/vox1/test/wav'
Gaël Le Lan's avatar
Gaël Le Lan committed
236

Gaël Le Lan's avatar
Gaël Le Lan committed
237
238
239
    transform_pipeline = dict()

    xv_stat = extract_embeddings(idmap_name=idmap_test_filename,
Gaël Le Lan's avatar
Gaël Le Lan committed
240
                                 model_filename=model,
Gaël Le Lan's avatar
Gaël Le Lan committed
241
                                 data_root_name=data_root_name,
Gaël Le Lan's avatar
Gaël Le Lan committed
242
                                 device=device,
Gaël Le Lan's avatar
Gaël Le Lan committed
243
244
                                 transform_pipeline=transform_pipeline,
                                 num_thread=num_thread,
Gaël Le Lan's avatar
Gaël Le Lan committed
245
246
                                 mixed_precision=mixed_precision,
                                 backward=False)
Gaël Le Lan's avatar
Gaël Le Lan committed
247

Gaël Le Lan's avatar
Gaël Le Lan committed
248
249
250
251
252
253
254
255
256
257
    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))

    #test_eer = eer(numpy.array(non).astype(numpy.double), numpy.array(tar).astype(numpy.double))
    pmiss, pfa = rocch(tar, non)
Gaël Le Lan's avatar
Gaël Le Lan committed
258

Gaël Le Lan's avatar
Gaël Le Lan committed
259
    return rocch2eer(pmiss, pfa)
Gaël Le Lan's avatar
Gaël Le Lan committed
260

Anthony Larcher's avatar
Anthony Larcher committed
261

Anthony Larcher's avatar
Anthony Larcher committed
262
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
263
264
265
266
267
268
269
270
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
271
272
273
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
274

Anthony Larcher's avatar
Anthony Larcher committed
275

Anthony Larcher's avatar
Anthony Larcher committed
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
class MeanStdPooling(torch.nn.Module):
    """
    Mean and Standard deviation pooling
    """
    def __init__(self):
        """

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

    def forward(self, x):
        """

        :param x:
        :return:
        """
Gaël Le Lan's avatar
Gaël Le Lan committed
293

Anthony Larcher's avatar
Anthony Larcher committed
294
295
296
        mean = torch.mean(x, dim=2)
        std = torch.std(x, dim=2)
        return torch.cat([mean, std], dim=1)
Anthony Larcher's avatar
Anthony Larcher committed
297

Anthony Larcher's avatar
Anthony Larcher committed
298

Gaël Le Lan's avatar
Gaël Le Lan committed
299
300
301
302
class AttentivePooling(torch.nn.Module):
    """
    Mean and Standard deviation attentive pooling
    """
Gaël Le Lan's avatar
Gaël Le Lan committed
303
    def __init__(self, num_channels, n_mels):
Gaël Le Lan's avatar
Gaël Le Lan committed
304
305
306
        """

        """
Gaël Le Lan's avatar
Gaël Le Lan committed
307
308
        # TODO Make global_context configurable (True/False)
        # TODO Make convolution parameters configurable
Gaël Le Lan's avatar
Gaël Le Lan committed
309
        super(AttentivePooling, self).__init__()
Gaël Le Lan's avatar
Gaël Le Lan committed
310
        self.attention = torch.nn.Sequential(
Gaël Le Lan's avatar
Gaël Le Lan committed
311
            torch.nn.Conv1d(num_channels * (n_mels//8), num_channels//32, kernel_size=1),
Gaël Le Lan's avatar
Gaël Le Lan committed
312
            torch.nn.ReLU(),
Gaël Le Lan's avatar
Gaël Le Lan committed
313
            torch.nn.BatchNorm1d(num_channels//32),
Gaël Le Lan's avatar
Gaël Le Lan committed
314
            torch.nn.Tanh(),
Gaël Le Lan's avatar
Gaël Le Lan committed
315
            torch.nn.Conv1d(num_channels//32, num_channels * (n_mels//8), kernel_size=1),
Gaël Le Lan's avatar
Gaël Le Lan committed
316
317
            torch.nn.Softmax(dim=2),
            )
Gaël Le Lan's avatar
Gaël Le Lan committed
318
        #self.global_context = MeanStdPooling()
Gaël Le Lan's avatar
Gaël Le Lan committed
319
320
321
322
323
324
325
326
327
328
329
330

    def new_parameter(self, *size):
        out = torch.nn.Parameter(torch.FloatTensor(*size))
        torch.nn.init.xavier_normal_(out)
        return out

    def forward(self, x):
        """

        :param x:
        :return:
        """
Gaël Le Lan's avatar
Gaël Le Lan committed
331

Gaël Le Lan's avatar
Gaël Le Lan committed
332
        #global_context = self.global_context(x).unsqueeze(2).repeat(1, 1, x.shape[-1])
Gaël Le Lan's avatar
Gaël Le Lan committed
333

Gaël Le Lan's avatar
Gaël Le Lan committed
334
335
        #w = self.attention(torch.cat([x, global_context], dim=1))
        w = self.attention(x)
Gaël Le Lan's avatar
Gaël Le Lan committed
336
337
338
339
340

        mu = torch.sum(x * w, dim=2)
        rh = torch.sqrt( ( torch.sum((x**2) * w, dim=2) - mu**2 ).clamp(min=1e-5) )
        x = torch.cat((mu, rh),1)
        x = x.view(x.size()[0], -1)
Gaël Le Lan's avatar
Gaël Le Lan committed
341
342
343
        return x


Anthony Larcher's avatar
Anthony Larcher committed
344
345
346
347
348
349
350
351
352
353
354
class GruPooling(torch.nn.Module):
    """

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

        :param input_size:
        :param gru_node:
        :param nb_gru_layer:
        """
Anthony Larcher's avatar
Anthony Larcher committed
355
        super(GruPooling, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
356
357
358
        self.lrelu_keras = torch.nn.LeakyReLU(negative_slope = 0.3)
        self.bn_before_gru = torch.nn.BatchNorm1d(num_features = input_size)
        self.gru = torch.nn.GRU(input_size = input_size,
Anthony Larcher's avatar
Anthony Larcher committed
359
360
361
                                hidden_size = gru_node,
                                num_layers = nb_gru_layer,
                                batch_first = True)
Anthony Larcher's avatar
Anthony Larcher committed
362
363
364
365
366
367
368
369
370
371
372
373

    def forward(self, x):
        """

        :param x:
        :return:
        """
        x = self.bn_before_gru(x)
        x = self.lrelu_keras(x)
        x = x.permute(0, 2, 1)  #(batch, filt, time) >> (batch, time, filt)
        self.gru.flatten_parameters()
        x, _ = self.gru(x)
Anthony Larcher's avatar
Anthony Larcher committed
374
        x = x[:, -1, :]
Anthony Larcher's avatar
Anthony Larcher committed
375
376
377

        return x

Anthony Larcher's avatar
Anthony Larcher committed
378

Anthony Larcher's avatar
Anthony Larcher committed
379
class Xtractor(torch.nn.Module):
380
381
382
    """
    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
383

Anthony Larcher's avatar
Anthony Larcher committed
384
385
386
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
387
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
388
                 norm_embedding=False,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
389
390
                 aam_margin=0.2,
                 aam_s=30):
Anthony Larcher's avatar
Anthony Larcher committed
391
392
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
393
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
394
        """
Anthony Larcher's avatar
Anthony Larcher committed
395
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
396
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
397
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
398
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
399

Anthony Larcher's avatar
Anthony Larcher committed
400
401
        print(f"Speaker number : {self.speaker_number}")

Anthony Larcher's avatar
Anthony Larcher committed
402
        if model_archi == "xvector":
403
404
405
406
407
408
409

            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
410

Anthony Larcher's avatar
Anthony Larcher committed
411
412
            self.activation = torch.nn.LeakyReLU(0.2)

Gaël Le Lan's avatar
Gaël Le Lan committed
413
414
            self.preprocessor = MfccFrontEnd()
            self.feature_size = self.preprocessor.n_mfcc
Anthony Larcher's avatar
Anthony Larcher committed
415

Anthony Larcher's avatar
xv    
Anthony Larcher committed
416
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
417
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
418
                ("activation1", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
419
                ("batch_norm1", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
420
421
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
422
                ("batch_norm2", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
423
424
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
425
                ("batch_norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
426
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
427
                ("activation4", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
428
                ("batch_norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
429
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
430
                ("activation5", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
431
                ("batch_norm5", torch.nn.BatchNorm1d(1536))
Anthony Larcher's avatar
Anthony Larcher committed
432
433
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
434
            self.stat_pooling = MeanStdPooling()
Gaël Le Lan's avatar
Gaël Le Lan committed
435
            self.stat_pooling_weight_decay = 0
Anthony Larcher's avatar
xv    
Anthony Larcher committed
436
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Gaël Le Lan's avatar
Gaël Le Lan committed
437
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
438
439
            ]))

440
441
            self.embedding_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
442
            if self.loss == "aam":
Gaël Le Lan's avatar
Gaël Le Lan committed
443
                self.after_speaker_embedding = ArcMarginProduct(512,
Anthony Larcher's avatar
Anthony Larcher committed
444
445
446
447
                                                                int(self.speaker_number),
                                                                s=64,
                                                                m=0.2,
                                                                easy_margin=True)
Gaël Le Lan's avatar
Gaël Le Lan committed
448
            elif self.loss == "cce":
Gaël Le Lan's avatar
Gaël Le Lan committed
449
                self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
Gaël Le Lan's avatar
Gaël Le Lan committed
450
                    ("activation6", torch.nn.LeakyReLU(0.2)),
Gaël Le Lan's avatar
Gaël Le Lan committed
451
                    ("batch_norm6", torch.nn.BatchNorm1d(512)),
Gaël Le Lan's avatar
Gaël Le Lan committed
452
                    ("dropout6", torch.nn.Dropout(p=0.05)),
Gaël Le Lan's avatar
Gaël Le Lan committed
453
                    ("linear7", torch.nn.Linear(512, 512)),
Gaël Le Lan's avatar
Gaël Le Lan committed
454
                    ("activation7", torch.nn.LeakyReLU(0.2)),
Gaël Le Lan's avatar
Gaël Le Lan committed
455
456
                    ("batch_norm7", torch.nn.BatchNorm1d(512)),
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
Gaël Le Lan's avatar
Gaël Le Lan committed
457
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
458

Gaël Le Lan's avatar
Gaël Le Lan committed
459
            self.preprocessor_weight_decay = 0.0002
460
461
462
463
464
465
            self.sequence_network_weight_decay = 0.0002
            self.before_speaker_embedding_weight_decay = 0.002
            self.after_speaker_embedding_weight_decay = 0.002
            self.embedding_size = 512

        elif model_archi == "resnet34":
Gaël Le Lan's avatar
Gaël Le Lan committed
466
467

            self.preprocessor = MfccFrontEnd()
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
            self.sequence_network = PreResNet34()

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

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

            self.embedding_size = 256

            self.loss = "aam"
            self.after_speaker_embedding = ArcMarginProduct(256,
                                                            int(self.speaker_number),
                                                            s = 30.0,
                                                            m = 0.20,
                                                            easy_margin = True)

            self.preprocessor_weight_decay = 0.000
Gaël Le Lan's avatar
Gaël Le Lan committed
486
            self.sequence_network_weight_decay = 0.000
487
            self.stat_pooling_weight_decay = 0.000
Gaël Le Lan's avatar
Gaël Le Lan committed
488
489
            self.before_speaker_embedding_weight_decay = 0.00
            self.after_speaker_embedding_weight_decay = 0.00
Anthony Larcher's avatar
Anthony Larcher committed
490

491
        elif model_archi == "fastresnet34":
Gaël Le Lan's avatar
Gaël Le Lan committed
492
            self.preprocessor = MelSpecFrontEnd()
Gaël Le Lan's avatar
Gaël Le Lan committed
493
            self.sequence_network = PreFastResNet34()
Gaël Le Lan's avatar
Gaël Le Lan committed
494
            self.embedding_size = 256
Gaël Le Lan's avatar
Gaël Le Lan committed
495

Gaël Le Lan's avatar
Gaël Le Lan committed
496
            self.before_speaker_embedding = torch.nn.Linear(in_features = 2560,
Gaël Le Lan's avatar
Gaël Le Lan committed
497
                                                            out_features = self.embedding_size)
Anthony Larcher's avatar
Anthony Larcher committed
498

Gaël Le Lan's avatar
Gaël Le Lan committed
499
            self.stat_pooling = AttentivePooling(128)
Anthony Larcher's avatar
Anthony Larcher committed
500
501
            self.stat_pooling_weight_decay = 0

Gaël Le Lan's avatar
Gaël Le Lan committed
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
            self.loss = loss
            if self.loss == "aam":
                self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
                                                                int(self.speaker_number),
                                                                s = 30,
                                                                m = 0.2,
                                                                easy_margin = False)
            
            elif self.loss == 'aps':
                self.after_speaker_embedding = SoftmaxAngularProto(int(self.speaker_number))

            self.preprocessor_weight_decay = 0.000
            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
        elif model_archi == "halfresnet34":
Gaël Le Lan's avatar
Gaël Le Lan committed
519
520
            self.preprocessor = MelSpecFrontEnd(n_fft=512, win_length=400, hop_length=160, n_mels=64)
            #self.preprocessor = MelSpecFrontEnd()
Gaël Le Lan's avatar
Gaël Le Lan committed
521
522
523
            self.sequence_network = PreHalfResNet34()
            self.embedding_size = 512

Gaël Le Lan's avatar
Gaël Le Lan committed
524
            self.before_speaker_embedding = torch.nn.Linear(in_features = 4096,
Gaël Le Lan's avatar
Gaël Le Lan committed
525
526
                                                            out_features = self.embedding_size)

Gaël Le Lan's avatar
Gaël Le Lan committed
527
            self.stat_pooling = AttentivePooling(256, 64)
Gaël Le Lan's avatar
Gaël Le Lan committed
528
529
            self.stat_pooling_weight_decay = 0

Gaël Le Lan's avatar
Gaël Le Lan committed
530
531
532
533
534
535
536
537
538
539
            self.loss = loss
            if self.loss == "aam":
                self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
                                                                int(self.speaker_number),
                                                                s = 30,
                                                                m = 0.2,
                                                                easy_margin = False)
            
            elif self.loss == 'aps':
                self.after_speaker_embedding = SoftmaxAngularProto(int(self.speaker_number))
Anthony Larcher's avatar
Anthony Larcher committed
540
541
542
543
544
545
546

            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
547
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
548
549
550
551
552
553

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

554
555
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
556
557
558
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
559
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
560
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
561
562
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579

            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
580
581
582
583
584
            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
585
            elif self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
586
587
588
                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
589

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
590
591
592
593
594
595
            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
596
        else:
Anthony Larcher's avatar
Anthony Larcher committed
597
598
            is_first_resblock = True

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
599
600
601
602
603
604
            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
605

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
606
            self.loss = cfg["training"]["loss"]
Anthony Larcher's avatar
Anthony Larcher committed
607
            if self.loss == "aam":
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
608
609
                self.aam_margin = cfg["training"]["aam_margin"]
                self.aam_s = cfg["training"]["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
610

Anthony Larcher's avatar
Anthony Larcher committed
611
612
613
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
614
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
615
616
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
617
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
618
619
620
621
622
623
624
625
626
627
628
629
                        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
630
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
631
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
632
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
633
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
634
635
636
637
638
639
                                                        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
640
                self.preprocessor_weight_decay = 0.000
Anthony Larcher's avatar
Anthony Larcher committed
641
642

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
643
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
644
            """
Anthony Larcher's avatar
Anthony Larcher committed
645
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
646
647
648
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
            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
664
665
666
667
668
                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
669
670
671
672
673
                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
674
                                                                stride=1)))
Anthony Larcher's avatar
Anthony Larcher committed
675

Anthony Larcher's avatar
Anthony Larcher committed
676
                elif k.startswith("conv"):
Anthony Larcher's avatar
Anthony Larcher committed
677
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
678
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
679
680
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
681
682
                    input_size = cfg["segmental"][k]["output_channels"]

Anthony Larcher's avatar
Anthony Larcher committed
683
684
                elif k.startswith("ctrans"):
                    segmental_layers.append((k, torch.nn.ConvTranspose1d(input_size,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
685
                                                                         cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
686
687
                                                                         kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                         dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
688
689
690
                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
691
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
Anthony Larcher committed
692
693
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
694
695
696
697
698
699
                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
700
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
701
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
702

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
703
704
705
706
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
707
            tmp_input_size = input_size * 2
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
708
709
710
711
            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
712
                tmp_input_size = cfg["stat_pooling"]["gru_node"]
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
713

Anthony Larcher's avatar
Anthony Larcher committed
714
715
            self.stat_pooling_weight_decay = cfg["stat_pooling"]["weight_decay"]

Anthony Larcher's avatar
Anthony Larcher committed
716
            """
Anthony Larcher's avatar
Anthony Larcher committed
717
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
718
            """
Anthony Larcher's avatar
Anthony Larcher committed
719
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
720
            input_size = tmp_input_size
Anthony Larcher's avatar
xv    
Anthony Larcher committed
721
722
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
723
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
724
725
                    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
726
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
727
728
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
729
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
730
731

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
732
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
733

Anthony Larcher's avatar
Anthony Larcher committed
734
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
735
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
736
737

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
738
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
739

Anthony Larcher's avatar
Anthony Larcher committed
740
            self.embedding_size = input_size
Anthony Larcher's avatar
Anthony Larcher committed
741
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
742
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
743

Anthony Larcher's avatar
Anthony Larcher committed
744
            # if loss_criteria is "cce"
Anthony Larcher's avatar
xv    
Anthony Larcher committed
745
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
746
747
748
749
750
751
752
753
            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
754
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
755
                            input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
756

Anthony Larcher's avatar
Anthony Larcher committed
757
758
759
760
761
                    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
762

Anthony Larcher's avatar
Anthony Larcher committed
763
764
                    elif k.startswith("activation"):
                        after_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
765

Anthony Larcher's avatar
Anthony Larcher committed
766
767
                    elif k.startswith('batch_norm'):
                        after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
768

Anthony Larcher's avatar
Anthony Larcher committed
769
770
771
772
773
774
                    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
775
                self.norm_embedding = True
Anthony Larcher's avatar
Anthony Larcher committed
776
777
778
779
780
781
                self.after_speaker_embedding = ArcMarginProduct(input_size,
                                                                int(self.speaker_number),
                                                                s=64,
                                                                m=0.2,
                                                                easy_margin=True)

Anthony Larcher's avatar
arcface    
Anthony Larcher committed
782
783
784
785
                #self.after_speaker_embedding = ArcLinear(input_size,
                #                                         self.speaker_number,
                #                                         margin=self.aam_margin,
                #                                         s=self.aam_s)
Anthony Larcher's avatar
Anthony Larcher committed
786
787
788
789
                #self.after_speaker_embedding = ArcFace(embedding_size=input_size,
                #                                       classnum=self.speaker_number,
                #                                       s=64.,
                #                                       m=0.5)
Anthony Larcher's avatar
Anthony Larcher committed
790

Anthony Larcher's avatar
Anthony Larcher committed
791
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
792

Anthony Larcher's avatar
Anthony Larcher committed
793

794
    def forward(self, x, is_eval=False, target=None, extract_after_pooling=False):
795
796
797
        """

        :param x:
798
        :param is_eval: False for training
799
800
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
801
802
803
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
804
        x = self.sequence_network(x)
805

Anthony Larcher's avatar
Anthony Larcher committed
806
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
807
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
808

809
810
        if extract_after_pooling:
            return x
811

Anthony Larcher's avatar
Anthony Larcher committed
812
        x = self.before_speaker_embedding(x)
813

Anthony Larcher's avatar
Anthony Larcher committed
814
        if self.norm_embedding:
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
815
            #x_norm = x.norm(p=2,dim=1, keepdim=True) / 10. # Why  10. ?
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
816
817
818
            #x_norm = torch.linalg.norm(x, ord=2, dim=1, keepdim=True, out=None, dtype=None)
            #x = torch.div(x, x_norm)
            x = l2_norm(x)
Gaël Le Lan's avatar
Gaël Le Lan committed
819

Anthony Larcher's avatar
Anthony Larcher committed
820
        if self.loss == "cce":
Gaël Le Lan's avatar
Gaël Le Lan committed
821
            if is_eval:
Gaël Le Lan's avatar
Gaël Le Lan committed
822
                return x
Gaël Le Lan's avatar
Gaël Le Lan committed
823
            else:
Gaël Le Lan's avatar
Gaël Le Lan committed
824
                return self.after_speaker_embedding(x), x
Anthony Larcher's avatar
Anthony Larcher committed
825

Gaël Le Lan's avatar
Gaël Le Lan committed
826
        elif self.loss in ['aam', 'aps']:
Gaël Le Lan's avatar
Gaël Le Lan committed
827
828
            if is_eval:
                x = torch.nn.functional.normalize(x, dim=1)
Anthony Larcher's avatar
Anthony Larcher committed
829
            else:
Gaël Le Lan's avatar
Gaël Le Lan committed
830
                x = self.after_speaker_embedding(x, target=target), torch.nn.functional.normalize(x, dim=1)
Gaël Le Lan's avatar
Gaël Le Lan committed
831

Anthony Larcher's avatar
Anthony Larcher committed
832
        return x
Anthony Larcher's avatar
Anthony Larcher committed
833

834
835
    def context_size(self):
        context = 1
Anthony Larcher's avatar
Anthony Larcher committed
836
837
838
839
840
841
842
843
844
        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
845

Anthony Larcher's avatar
Anthony Larcher committed
846
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
847
           dataset_yaml,
Anthony Larcher's avatar
Anthony Larcher committed
848
849
           epochs=None,
           lr=None,
Anthony Larcher's avatar
Anthony Larcher committed
850
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
851
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
852
853
854
855
           loss=None,
           aam_margin=None,
           aam_s=None,
           patience=None,
Anthony Larcher's avatar
Anthony Larcher committed
856
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
857
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
858
           multi_gpu=True,
Gaël Le Lan's avatar
Gaël Le Lan committed
859
           device=None,
Gaël Le Lan's avatar
Gaël Le Lan committed
860
           mixed_precision=False,
861
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
862
           opt=None,
Anthony Larcher's avatar
Anthony Larcher committed
863
864
           reset_parts=[],
           freeze_parts=[],
Anthony Larcher's avatar
Anthony Larcher committed
865
           num_thread=None,
Gaël Le Lan's avatar
Gaël Le Lan committed
866
           compute_test_eer=True):
867
868
    """

Anthony Larcher's avatar
Anthony Larcher committed
869
870
871
872
873
874
    :param speaker_number:
    :param dataset_yaml:
    :param epochs:
    :param lr:
    :param model_yaml:
    :param model_name:
Anthony Larcher's avatar
Anthony Larcher committed
875
876
877
878
    :param loss:
    :param aam_margin:
    :param aam_s:
    :param patience:
Anthony Larcher's avatar
Anthony Larcher committed
879
880
881
    :param tmp_model_name:
    :param best_model_name:
    :param multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
882
    :param mixed_precision:
Anthony Larcher's avatar
Anthony Larcher committed
883
    :param clipping:
Anthony Larcher's avatar
Anthony Larcher committed
884
885
886
    :param opt:
    :param reset_parts:
    :param freeze_parts:
Anthony Larcher's avatar
Anthony Larcher committed
887
    :param num_thread:
Anthony Larcher's avatar
Anthony Larcher committed
888
    :param compute_test_eer:
889
890
    :return:
    """
Anthony Larcher's avatar
debug    
Anthony Larcher committed
891
892
    # Test to optimize
    torch.autograd.profiler.emit_nvtx(enabled=False)
Anthony Larcher's avatar
Anthony Larcher committed
893

Anthony Larcher's avatar
Anthony Larcher committed
894
    if num_thread is None:
Gaël Le Lan's avatar
Gaël Le Lan committed
895
        import multiprocessing
Anthony Larcher's avatar
Anthony Larcher committed
896
897
        num_thread = multiprocessing.cpu_count()

Anthony Larcher's avatar
Anthony Larcher committed
898
    logging.critical(f"Use {num_thread} cpus")
Anthony Larcher's avatar
Anthony Larcher committed
899
    logging.critical(f"Start process at {time.strftime('%H:%M:%S', time.localtime())}")
900

Gaël Le Lan's avatar
Gaël Le Lan committed
901
902
    if device == None:
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
903
904

    # Use a predefined architecture
Gaël Le Lan's avatar
Gaël Le Lan committed
905
    if model_yaml in ["xvector", "rawnet2", "resnet34", "fastresnet34", "halfresnet34"]:
906
907

        if model_name is None:
Gaël Le Lan's avatar
debug    
Gaël Le Lan committed
908
            model = Xtractor(speaker_number, model_yaml, loss=loss)
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931

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

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

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

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

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

Gaël Le Lan's avatar
Gaël Le Lan committed
932
        model_archi = model_yaml
933
934

    # Here use a config file to build the architecture
Anthony Larcher's avatar
Anthony Larcher committed
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
    else:
        with open(model_yaml, 'r') as fh:
            model_archi = yaml.load(fh, Loader=yaml.FullLoader)
            if epochs is None:
                epochs = model_archi["training"]["epochs"]
            if patience is None:
                patience = model_archi["training"]["patience"]
            if opt is None:
                opt = model_archi["training"]["opt"]
            if lr is None:
                lr = model_archi["training"]["lr"]
            if loss is None:
                loss = model_archi["training"]["loss"]
            if aam_margin is None and model_archi["training"]["loss"] == "aam":
                aam_margin = model_archi["training"]["aam_margin"]
            if aam_s is None and model_archi["training"]["loss"] == "aam":
                aam_s = model_archi["training"]["aam_s"]
            if tmp_model_name is None:
                tmp_model_name = model_archi["training"]["tmp_model_name"]
            if best_model_name is None:
                best_model_name = model_archi["training"]["best_model_name"]
            if multi_gpu is None:
                multi_gpu = model_archi["training"]["multi_gpu"]
            if clipping is None:
                clipping = model_archi["training"]["clipping"]

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
961
        if model_name is None:
Anthony Larcher's avatar
Anthony Larcher committed
962
963
964
            model = Xtractor(speaker_number, model_yaml)

         # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
965
        else:
Anthony Larcher's avatar
Anthony Larcher committed
966
967
            # Load the model
            logging.critical(f"*** Load model from = {model_name}")
Gaël Le Lan's avatar
Gaël Le Lan committed
968
969
            checkpoint = torch.load(model_name, map_location=device)
            model = Xtractor(speaker_number, model_yaml, loss=loss)
Anthony Larcher's avatar
Anthony Larcher committed
970

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
971
972
            """
            Here we remove all layers that we don't want to reload
Anthony Larcher's avatar
Anthony Larcher committed
973
        
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
974
975
976
977
978
979
980
981
982
983
984
985
986
            """
            pretrained_dict = checkpoint["model_state_dict"]
            for part in reset_parts:
                pretrained_dict = {k: v for k, v in pretrained_dict.items() if not k.startswith(part)}

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

        # Freeze required layers
        for name, param in model.named_parameters():
            if name.split(".")[0] in freeze_parts:
                param.requires_grad = False
Anthony Larcher's avatar
Anthony Larcher committed
987

Gaël Le Lan's avatar
Gaël Le Lan committed
988
989
990
991
    logging.critical(model)

    logging.critical("model_parameters_count: {:d}".format(
        sum(p.numel()
Gaël Le Lan's avatar
Gaël Le Lan committed
992
993
994
995
            for p in model.sequence_network.parameters()
            if p.requires_grad) + \
        sum(p.numel()
            for p in model.before_speaker_embedding.parameters()
Gaël Le Lan's avatar
Gaël Le Lan committed
996
997
998
            if p.requires_grad) + \
        sum(p.numel()
            for p in model.stat_pooling.parameters()
Gaël Le Lan's avatar
Gaël Le Lan committed
999
1000
            if p.requires_grad)))