xvector.py 77.9 KB
Newer Older
Anthony Larcher's avatar
Anthony Larcher committed
1
# coding: utf-8 -*-
Anthony Larcher's avatar
Anthony Larcher committed
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
#
# This file is part of SIDEKIT.
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is free software: you can redistribute it and/or modify
# it under the terms of the GNU LLesser General Public License as
# published by the Free Software Foundation, either version 3 of the License,
# or (at your option) any later version.
#
# SIDEKIT is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU Lesser General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with SIDEKIT.  If not, see <http://www.gnu.org/licenses/>.

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

Anthony Larcher's avatar
Anthony Larcher committed
28
29
30

import pdb
import traceback
Anthony Larcher's avatar
Anthony Larcher committed
31
import logging
Anthony Larcher's avatar
debug    
Anthony Larcher committed
32
import math
Anthony Larcher's avatar
Anthony Larcher committed
33
import os
Anthony Larcher's avatar
Anthony Larcher committed
34
import numpy
Anthony Larcher's avatar
Anthony Larcher committed
35
import pandas
Anthony Larcher's avatar
minor    
Anthony Larcher committed
36
import pickle
Anthony Larcher's avatar
Anthony Larcher committed
37
import shutil
Anthony Larcher's avatar
Anthony Larcher committed
38
import sys
39
import time
Anthony Larcher's avatar
Anthony Larcher committed
40
import torch
Anthony Larcher's avatar
debug    
Anthony Larcher committed
41
import torchaudio
Anthony Larcher's avatar
Anthony Larcher committed
42
import tqdm
Anthony Larcher's avatar
Anthony Larcher committed
43
44
import yaml

Anthony Larcher's avatar
Anthony Larcher committed
45
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
46
from .xsets import SideSet
Anthony Larcher's avatar
debug    
Anthony Larcher committed
47
from .xsets import FileSet
Anthony Larcher's avatar
Anthony Larcher committed
48
from .xsets import IdMapSet
Anthony Larcher's avatar
Anthony Larcher committed
49
from .xsets import IdMapSet_per_speaker
Anthony Larcher's avatar
debug    
Anthony Larcher committed
50
from .xsets import SpkSet
Anthony Larcher's avatar
Anthony Larcher committed
51
from .xsets import SideSampler
Anthony Larcher's avatar
Anthony Larcher committed
52
from .res_net import RawPreprocessor, ResBlockWFMS, ResBlock, PreResNet34, PreFastResNet34
Anthony Larcher's avatar
Anthony Larcher committed
53
from ..bosaris import IdMap
Anthony Larcher's avatar
Anthony Larcher committed
54
55
from ..bosaris import Key
from ..bosaris import Ndx
Anthony Larcher's avatar
Anthony Larcher committed
56
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
57
from ..iv_scoring import cosine_scoring
Anthony Larcher's avatar
Anthony Larcher committed
58
from torch.utils.data import DataLoader
Anthony Larcher's avatar
Anthony Larcher committed
59
from sklearn.model_selection import train_test_split
Anthony Larcher's avatar
Anthony Larcher committed
60
from .sincnet import SincNet
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
61
62
from .loss import ArcLinear
from .loss import l2_norm
Anthony Larcher's avatar
Anthony Larcher committed
63
from .loss import ArcMarginProduct
Anthony Larcher's avatar
Anthony Larcher committed
64

Anthony Larcher's avatar
ddp    
Anthony Larcher committed
65

Anthony Larcher's avatar
Anthony Larcher committed
66
os.environ['MKL_THREADING_LAYER'] = 'GNU'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
67

Anthony Larcher's avatar
Anthony Larcher committed
68
69
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
v1.3.7    
Anthony Larcher committed
70
__copyright__ = "Copyright 2015-2021 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
71
72
73
74
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
75
76


Anthony Larcher's avatar
Anthony Larcher committed
77
78
logging.basicConfig(format='%(asctime)s %(message)s')

Anthony Larcher's avatar
Anthony Larcher committed
79
80
81
82
83
84
85
86

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


Anthony Larcher's avatar
debug    
Anthony Larcher committed
87
88
def test_metrics(model,
                 device,
Anthony Larcher's avatar
Anthony Larcher committed
89
90
91
                 speaker_number,
                 num_thread,
                 mixed_precision):
Anthony Larcher's avatar
Anthony Larcher committed
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
    """Compute model metrics

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

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

    Returns:
        [type]: [description]
    """
Anthony Larcher's avatar
debug    
Anthony Larcher committed
108
109
110
    idmap_test_filename = 'h5f/idmap_test.h5'
    ndx_test_filename = 'h5f/ndx_test.h5'
    key_test_filename = 'h5f/key_test.h5'
Anthony Larcher's avatar
Anthony Larcher committed
111
    data_root_name='/data/larcher/voxceleb1/test/wav'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
112
113
114
115
116
117
118
119

    transform_pipeline = dict()

    xv_stat = extract_embeddings(idmap_name=idmap_test_filename,
                                 speaker_number=speaker_number,
                                 model_filename=model,
                                 data_root_name=data_root_name,
                                 device=device,
Anthony Larcher's avatar
Anthony Larcher committed
120
121
122
                                 transform_pipeline=transform_pipeline,
                                 num_thread=num_thread,
                                 mixed_precision=mixed_precision)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
123
124
125
126
127
128
129
130
131
132

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

    tar, non = scores.get_tar_non(Key(key_test_filename))
    test_eer = eer(numpy.array(non).astype(numpy.double), numpy.array(tar).astype(numpy.double))
    return test_eer
Anthony Larcher's avatar
Anthony Larcher committed
133

Anthony Larcher's avatar
Anthony Larcher committed
134

Anthony Larcher's avatar
Anthony Larcher committed
135
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
136
137
138
139
140
141
142
143
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
144
145
146
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
147

Anthony Larcher's avatar
Anthony Larcher committed
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
class MeanStdPooling(torch.nn.Module):
    """
    Mean and Standard deviation pooling
    """
    def __init__(self):
        """

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

    def forward(self, x):
        """

        :param x:
        :return:
        """
        mean = torch.mean(x, dim=2)
        std = torch.std(x, dim=2)
        return torch.cat([mean, std], dim=1)
Anthony Larcher's avatar
Anthony Larcher committed
168

Anthony Larcher's avatar
Anthony Larcher committed
169

Anthony Larcher's avatar
Anthony Larcher committed
170
171
172
173
174
175
176
177
178
179
180
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
181
        super(GruPooling, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
        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,
                          hidden_size = gru_node,
                          num_layers = nb_gru_layer,
                          batch_first = True)

    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)
        x = x[:,-1,:]

        return x

Anthony Larcher's avatar
Anthony Larcher committed
204

Anthony Larcher's avatar
Anthony Larcher committed
205
206
207
208
209
210
211
212
213

class MfccFrontEnd(torch.nn.Module):
    """

    """

    def __init__(self,
                 pre_emphasis=0.97,
                 sample_rate=16000,
Anthony Larcher's avatar
Anthony Larcher committed
214
                 n_fft=2048,
Anthony Larcher's avatar
Anthony Larcher committed
215
216
217
218
                 f_min=133.333,
                 f_max=6855.4976,
                 win_length=1024,
                 window_fn=torch.hann_window,
Anthony Larcher's avatar
Anthony Larcher committed
219
                 hop_length=512,
Anthony Larcher's avatar
Anthony Larcher committed
220
                 power=2.0,
Anthony Larcher's avatar
Anthony Larcher committed
221
                 n_mels=100,
Anthony Larcher's avatar
Anthony Larcher committed
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
                 n_mfcc=80):

        super(MfccFrontEnd, self).__init__()

        self.pre_emphasis = pre_emphasis
        self.sample_rate = sample_rate
        self.n_fft = n_fft
        self.f_min = f_min
        self.f_max = f_max
        self.win_length = win_length
        self.window_fn=window_fn
        self.hop_length = hop_length
        self.power=power
        self.window_fn = window_fn
        self.n_mels = n_mels
        self.n_mfcc = n_mfcc

        self.PreEmphasis = PreEmphasis(self.pre_emphasis)

Anthony Larcher's avatar
Anthony Larcher committed
241
242
243
244
245
246
247
248
        self.melkwargs = {"n_fft":self.n_fft,
                          "f_min":self.f_min,
                          "f_max":self.f_max,
                          "win_length":self.win_length,
                          "window_fn":self.window_fn,
                          "hop_length":self.hop_length,
                          "power":self.power,
                          "n_mels":self.n_mels}
Anthony Larcher's avatar
Anthony Larcher committed
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271

        self.MFCC = torchaudio.transforms.MFCC(
            sample_rate=self.sample_rate,
            n_mfcc=self.n_mfcc,
            dct_type=2,
            log_mels=True,
            melkwargs=self.melkwargs)

        self.CMVN = torch.nn.InstanceNorm1d(self.n_mfcc)

    def forward(self, x):
        """

        :param x:
        :return:
        """
        with torch.no_grad():
            with torch.cuda.amp.autocast(enabled=False):
                mfcc = self.PreEmphasis(x)
                mfcc = self.MFCC(mfcc)
                mfcc = self.CMVN(mfcc)
        return mfcc

Anthony Larcher's avatar
FB    
Anthony Larcher committed
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341

class MelSpecFrontEnd(torch.nn.Module):
    """

    """

    def __init__(self,
                 pre_emphasis=0.97,
                 sample_rate=16000,
                 n_fft=1024,
                 f_min=90,
                 f_max=7600,
                 win_length=1024,
                 window_fn=torch.hann_window,
                 hop_length=256,
                 power=2.0,
                 n_mels=80):

        super(MelSpecFrontEnd, self).__init__()

        self.pre_emphasis = pre_emphasis
        self.sample_rate = sample_rate
        self.n_fft = n_fft
        self.f_min = f_min
        self.f_max = f_max
        self.win_length = win_length
        self.window_fn=window_fn
        self.hop_length = hop_length
        self.power=power
        self.window_fn = window_fn
        self.n_mels = n_mels

        self.PreEmphasis = PreEmphasis(self.pre_emphasis)

        self.melkwargs = {"n_fft":self.n_fft,
                          "f_min":self.f_min,
                          "f_max":self.f_max,
                          "win_length":self.win_length,
                          "window_fn":self.window_fn,
                          "hop_length":self.hop_length,
                          "power":self.power,
                          "n_mels":self.n_mels}

        self.MelSpec = torchaudio.transforms.MelSpectrogram(sample_rate=self.sample_rate,
                                                            n_fft=self.melkwargs['n_fft'],
                                                            f_min=self.melkwargs['f_min'],
                                                            f_max=self.melkwargs['f_max'],
                                                            win_length=self.melkwargs['win_length'],
                                                            hop_length=self.melkwargs['hop_length'],
                                                            window_fn=self.melkwargs['window_fn'],
                                                            power=self.melkwargs['power'],
                                                            n_mels=self.melkwargs['n_mels'])

        self.CMVN = torch.nn.InstanceNorm1d(self.n_mels)

    def forward(self, x):
        """

        :param x:
        :return:
        """
        with torch.no_grad():
            with torch.cuda.amp.autocast(enabled=False):
                out = self.PreEmphasis(x)
                out = self.MelSpec(out)+1e-6
                out = torch.log(out)
                out = self.CMVN(out)
        return out


Anthony Larcher's avatar
Anthony Larcher committed
342
class Xtractor(torch.nn.Module):
343
344
345
    """
    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
346

Anthony Larcher's avatar
Anthony Larcher committed
347
348
349
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
350
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
351
                 norm_embedding=False,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
352
353
                 aam_margin=0.2,
                 aam_s=30):
Anthony Larcher's avatar
Anthony Larcher committed
354
355
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
356
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
357
        """
Anthony Larcher's avatar
Anthony Larcher committed
358
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
359
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
360
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
361
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
362

Anthony Larcher's avatar
Anthony Larcher committed
363
364
        print(f"Speaker number : {self.speaker_number}")

Anthony Larcher's avatar
Anthony Larcher committed
365
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
366

Anthony Larcher's avatar
Anthony Larcher committed
367
368
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
369
370
371
372
373
            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
374
375
            self.activation = torch.nn.LeakyReLU(0.2)

Anthony Larcher's avatar
Anthony Larcher committed
376
            self.preprocessor = MfccFrontEnd()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
377
            self.feature_size = self.preprocessor.n_mfcc
Anthony Larcher's avatar
Anthony Larcher committed
378

Anthony Larcher's avatar
xv    
Anthony Larcher committed
379
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
380
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
381
                ("activation1", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
382
                ("batch_norm1", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
383
384
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
385
                ("batch_norm2", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
386
387
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
388
                ("batch_norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
389
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
390
                ("activation4", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
391
                ("batch_norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
392
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
393
                ("activation5", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
394
                ("batch_norm5", torch.nn.BatchNorm1d(1536))
Anthony Larcher's avatar
Anthony Larcher committed
395
396
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
397
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
398
            self.stat_pooling_weight_decay = 0
Anthony Larcher's avatar
xv    
Anthony Larcher committed
399
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
400
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
401
402
            ]))

403
404
            self.embedding_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
405
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
406
407
408
409
410
                self.after_speaker_embedding = ArcMarginProduct(512,
                                                                int(self.speaker_number),
                                                                s=64,
                                                                m=0.2,
                                                                easy_margin=True)
Anthony Larcher's avatar
Anthony Larcher committed
411
412
413
            elif self.loss == "cce":
                self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
                    ("activation6", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
414
                    ("batch_norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
415
416
417
                    ("dropout6", torch.nn.Dropout(p=0.05)),
                    ("linear7", torch.nn.Linear(512, 512)),
                    ("activation7", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
418
                    ("batch_norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
419
420
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
421

Anthony Larcher's avatar
debug    
Anthony Larcher committed
422
423
424
            self.sequence_network_weight_decay = 0.0002
            self.before_speaker_embedding_weight_decay = 0.002
            self.after_speaker_embedding_weight_decay = 0.002
425
            self.embedding_size = 512
Anthony Larcher's avatar
Anthony Larcher committed
426

Anthony Larcher's avatar
Anthony Larcher committed
427
        elif model_archi == "resnet34":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
428

Anthony Larcher's avatar
Anthony Larcher committed
429
            self.preprocessor = MfccFrontEnd()
Anthony Larcher's avatar
Anthony Larcher committed
430
431
432
433
434
435
436
437
            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

438
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
439

440
            self.loss = "aam"
Anthony Larcher's avatar
Anthony Larcher committed
441
442
443
            self.after_speaker_embedding = ArcMarginProduct(256,
                                                            int(self.speaker_number),
                                                            s = 30.0,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
444
                                                            m = 0.20,
Anthony Larcher's avatar
Anthony Larcher committed
445
                                                            easy_margin = True)
Anthony Larcher's avatar
Anthony Larcher committed
446
447
448
449
450
451
452

            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
453
        elif model_archi == "fastresnet34":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
454

Anthony Larcher's avatar
FB    
Anthony Larcher committed
455
            self.preprocessor = MelSpecFrontEnd()
Anthony Larcher's avatar
Anthony Larcher committed
456
            self.sequence_network = PreFastResNet34()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
457
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
458

Anthony Larcher's avatar
Anthony Larcher committed
459
            self.before_speaker_embedding = torch.nn.Linear(in_features = 2560,
Anthony Larcher's avatar
Anthony Larcher committed
460
                                                            out_features = self.embedding_size)
Anthony Larcher's avatar
Anthony Larcher committed
461
462
463
464
465

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

            self.loss = "aam"
Anthony Larcher's avatar
debug    
Anthony Larcher committed
466
            self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
Anthony Larcher's avatar
Anthony Larcher committed
467
                                                            int(self.speaker_number),
Anthony Larcher's avatar
debug    
Anthony Larcher committed
468
469
                                                            s = 30,
                                                            m = 0.2,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
470
                                                            easy_margin = False)
Anthony Larcher's avatar
Anthony Larcher committed
471
472
473
474
475
476

            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
477

Anthony Larcher's avatar
Anthony Larcher committed
478
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
479
480
481
482
483
484

            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
485
486
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
487
488
489
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
490
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
491
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
492
493
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510

            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
511
512
513
514
515
            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
516
            elif self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
517
518
519
                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
520

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
521
522
523
524
525
526
            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
527
        else:
Anthony Larcher's avatar
Anthony Larcher committed
528
529
            is_first_resblock = True

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
530
531
532
533
534
535
            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
536

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
537
            self.loss = cfg["training"]["loss"]
Anthony Larcher's avatar
Anthony Larcher committed
538
            if self.loss == "aam":
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
539
540
                self.aam_margin = cfg["training"]["aam_margin"]
                self.aam_s = cfg["training"]["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
541

Anthony Larcher's avatar
Anthony Larcher committed
542
543
544
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
545
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
546
547
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
548
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
549
550
551
552
553
554
555
556
557
558
559
560
                        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
561
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
562
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
563
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
564
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
565
566
567
568
569
570
                                                        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
571
                self.preprocessor_weight_decay = 0.000
Anthony Larcher's avatar
Anthony Larcher committed
572
573

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
574
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
575
            """
Anthony Larcher's avatar
Anthony Larcher committed
576
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
577
578
579
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
            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
595
596
597
598
599
                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
600
601
602
603
604
                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
605
                                                                stride=1)))
Anthony Larcher's avatar
Anthony Larcher committed
606

Anthony Larcher's avatar
Anthony Larcher committed
607
                elif k.startswith("conv"):
Anthony Larcher's avatar
Anthony Larcher committed
608
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
609
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
610
611
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
612
613
                    input_size = cfg["segmental"][k]["output_channels"]

Anthony Larcher's avatar
Anthony Larcher committed
614
615
                elif k.startswith("ctrans"):
                    segmental_layers.append((k, torch.nn.ConvTranspose1d(input_size,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
616
                                                                         cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
617
618
                                                                         kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                         dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
619
620
621
                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
622
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
Anthony Larcher committed
623
624
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
625
626
627
628
629
630
                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
631
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
632
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
633

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
634
635
636
637
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
638
            tmp_input_size = input_size * 2
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
639
640
641
642
            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
643
                tmp_input_size = cfg["stat_pooling"]["gru_node"]
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
644

Anthony Larcher's avatar
Anthony Larcher committed
645
646
            self.stat_pooling_weight_decay = cfg["stat_pooling"]["weight_decay"]

Anthony Larcher's avatar
Anthony Larcher committed
647
            """
Anthony Larcher's avatar
Anthony Larcher committed
648
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
649
            """
Anthony Larcher's avatar
Anthony Larcher committed
650
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
651
            input_size = tmp_input_size
Anthony Larcher's avatar
xv    
Anthony Larcher committed
652
653
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
654
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
655
656
                    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
657
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
658
659
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
660
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
661
662

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
663
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
664

Anthony Larcher's avatar
Anthony Larcher committed
665
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
666
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
667
668

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
669
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
670

Anthony Larcher's avatar
Anthony Larcher committed
671
            self.embedding_size = input_size
Anthony Larcher's avatar
Anthony Larcher committed
672
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
673
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
674

Anthony Larcher's avatar
Anthony Larcher committed
675
            # if loss_criteria is "cce"
Anthony Larcher's avatar
xv    
Anthony Larcher committed
676
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
677
678
679
680
681
682
683
684
            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
685
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
686
                            input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
687

Anthony Larcher's avatar
Anthony Larcher committed
688
689
690
691
692
                    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
693

Anthony Larcher's avatar
Anthony Larcher committed
694
695
                    elif k.startswith("activation"):
                        after_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
696

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

Anthony Larcher's avatar
Anthony Larcher committed
700
701
702
703
704
705
                    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
706
707
708
709
710
711
712
                self.norm_embedding = True
                self.after_speaker_embedding = ArcMarginProduct(input_size,
                                                                int(self.speaker_number),
                                                                s=64,
                                                                m=0.2,
                                                                easy_margin=True)

Anthony Larcher's avatar
arcface    
Anthony Larcher committed
713
714
715
716
                #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
717
718
719
720
                #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
721

Anthony Larcher's avatar
Anthony Larcher committed
722
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
723

Anthony Larcher's avatar
Anthony Larcher committed
724

725
    def forward(self, x, is_eval=False, target=None, extract_after_pooling=False):
726
727
728
        """

        :param x:
729
        :param is_eval: False for training
730
731
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
732
733
734
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
735
        x = self.sequence_network(x)
736

Anthony Larcher's avatar
Anthony Larcher committed
737
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
738
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
739

740
741
742
        if extract_after_pooling:
            return x

Anthony Larcher's avatar
Anthony Larcher committed
743
        x = self.before_speaker_embedding(x)
744

Anthony Larcher's avatar
Anthony Larcher committed
745
        if self.norm_embedding:
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
746
            #x_norm = x.norm(p=2,dim=1, keepdim=True) / 10. # Why  10. ?
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
747
748
749
            #x_norm = torch.linalg.norm(x, ord=2, dim=1, keepdim=True, out=None, dtype=None)
            #x = torch.div(x, x_norm)
            x = l2_norm(x)
Anthony Larcher's avatar
Anthony Larcher committed
750

Anthony Larcher's avatar
Anthony Larcher committed
751
        if self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
752
            if is_eval:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
753
                return x
Anthony Larcher's avatar
debug    
Anthony Larcher committed
754
            else:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
755
                return self.after_speaker_embedding(x), x
Anthony Larcher's avatar
Anthony Larcher committed
756

Anthony Larcher's avatar
Anthony Larcher committed
757
        elif self.loss == "aam":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
758
759
            if is_eval:
                x = torch.nn.functional.normalize(x, dim=1)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
760
            else:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
761
                x = self.after_speaker_embedding(torch.nn.functional.normalize(x, dim=1), target=target), torch.nn.functional.normalize(x, dim=1)
Anthony Larcher's avatar
Anthony Larcher committed
762

Anthony Larcher's avatar
Anthony Larcher committed
763
        return x
Anthony Larcher's avatar
Anthony Larcher committed
764

765
766
    def context_size(self):
        context = 1
Anthony Larcher's avatar
Anthony Larcher committed
767
768
769
770
771
772
773
774
775
        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
776

Anthony Larcher's avatar
Anthony Larcher committed
777
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
778
           dataset_yaml,
Anthony Larcher's avatar
Anthony Larcher committed
779
780
           epochs=None,
           lr=None,
Anthony Larcher's avatar
Anthony Larcher committed
781
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
782
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
783
784
785
786
           loss=None,
           aam_margin=None,
           aam_s=None,
           patience=None,
Anthony Larcher's avatar
Anthony Larcher committed
787
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
788
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
789
           multi_gpu=True,
Anthony Larcher's avatar
Anthony Larcher committed
790
           mixed_precision=False,
791
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
792
           opt=None,
Anthony Larcher's avatar
Anthony Larcher committed
793
794
           reset_parts=[],
           freeze_parts=[],
Anthony Larcher's avatar
Anthony Larcher committed
795
796
797
           num_thread=None,
           write_batches_to_disk=False,
           load_batches_from_disk=False,
Anthony Larcher's avatar
Anthony Larcher committed
798
799
           tmp_batch_dir=None,
           compute_test_eer=True):
800
801
    """

Anthony Larcher's avatar
Anthony Larcher committed
802
803
804
805
806
807
    :param speaker_number:
    :param dataset_yaml:
    :param epochs:
    :param lr:
    :param model_yaml:
    :param model_name:
Anthony Larcher's avatar
Anthony Larcher committed
808
809
810
811
    :param loss:
    :param aam_margin:
    :param aam_s:
    :param patience:
Anthony Larcher's avatar
Anthony Larcher committed
812
813
814
815
    :param tmp_model_name:
    :param best_model_name:
    :param multi_gpu:
    :param clipping:
Anthony Larcher's avatar
Anthony Larcher committed
816
817
818
    :param opt:
    :param reset_parts:
    :param freeze_parts:
Anthony Larcher's avatar
Anthony Larcher committed
819
    :param num_thread:
820
821
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
822
823
824
825
    # Add for tensorboard
    # Writer will output to ./runs/ directory by default
    #writer = SummaryWriter("runs/xvectors_experiments_2")
    writer = None
Anthony Larcher's avatar
Anthony Larcher committed
826

Anthony Larcher's avatar
debug    
Anthony Larcher committed
827
828
829
    # Test to optimize
    torch.autograd.profiler.emit_nvtx(enabled=False)

Anthony Larcher's avatar
Anthony Larcher committed
830
831
832
    if write_batches_to_disk:
        load_batches_from_disk = True

Anthony Larcher's avatar
Anthony Larcher committed
833
    if num_thread is None:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
834
        import multiprocessing
Anthony Larcher's avatar
Anthony Larcher committed
835
836
        num_thread = multiprocessing.cpu_count()

Anthony Larcher's avatar
Anthony Larcher committed
837
    logging.critical(f"Use {num_thread} cpus")
Anthony Larcher's avatar
Anthony Larcher committed
838
    logging.critical(f"Start process at {time.strftime('%H:%M:%S', time.localtime())}")
839

Anthony Larcher's avatar
debug    
Anthony Larcher committed
840
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869

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

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

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

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

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

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

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

Anthony Larcher's avatar
Anthony Larcher committed
870
        model_archi = model_yaml
871
872

    # Here use a config file to build the architecture
Anthony Larcher's avatar
Anthony Larcher committed
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
    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
899
        if model_name is None:
Anthony Larcher's avatar
Anthony Larcher committed
900
901
902
            model = Xtractor(speaker_number, model_yaml)

         # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
903
        else:
Anthony Larcher's avatar
Anthony Larcher committed
904
905
906
            # Load the model
            logging.critical(f"*** Load model from = {model_name}")
            checkpoint = torch.load(model_name)
Anthony Larcher's avatar
Anthony Larcher committed
907
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
908

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
909
910
            """
            Here we remove all layers that we don't want to reload
Anthony Larcher's avatar
Anthony Larcher committed
911
        
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
912
913
914
915
916
917
918
919
920
921
922
923
924
            """
            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
925

Anthony Larcher's avatar
Anthony Larcher committed
926
927
928
929
    logging.critical(model)

    logging.critical("model_parameters_count: {:d}".format(
        sum(p.numel()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
930
931
932
933
            for p in model.sequence_network.parameters()
            if p.requires_grad) + \
        sum(p.numel()
            for p in model.before_speaker_embedding.parameters()
Anthony Larcher's avatar
Anthony Larcher committed
934
935
            if p.requires_grad)))

Anthony Larcher's avatar
Anthony Larcher committed
936
    embedding_size = model.embedding_size
Anthony Larcher's avatar
Anthony Larcher committed
937

Anthony Larcher's avatar
Anthony Larcher committed
938
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
939
940
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
941
        #model = DDP(model)
Anthony Larcher's avatar
Anthony Larcher committed
942
943
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
944
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
945

Anthony Larcher's avatar
debug    
Anthony Larcher committed
946
947
948
949
    with open(dataset_yaml, "r") as fh:
        dataset_params = yaml.load(fh, Loader=yaml.FullLoader)
        df = pandas.read_csv(dataset_params["dataset_description"])

Anthony Larcher's avatar
Anthony Larcher committed
950
951
952
    if load_batches_from_disk:
        train_batch_fn_format = tmp_batch_dir + "/train/train_{}_batch.h5"
        val_batch_fn_format = tmp_batch_dir + "/val/val_{}_batch.h5"
953

Anthony Larcher's avatar
Anthony Larcher committed
954
955
956
957
958
959
960
    if not load_batches_from_disk or write_batches_to_disk:
        """
        Set the dataloaders according to the dataset_yaml
        
        First we load the dataframe from CSV file in order to split it for training and validation purpose
        Then we provide those two 
        """
Anthony Larcher's avatar
Anthony Larcher committed
961

Anthony Larcher's avatar
minor    
Anthony Larcher committed
962
        if write_batches_to_disk or dataset_params["batch_size"] > 1:
Anthony Larcher's avatar
Anthony Larcher committed
963
964
965
966
            output_format = "numpy"
        else:
            output_format = "pytorch"

Anthony Larcher's avatar
Anthony Larcher committed
967
968
        training_df, validation_df = train_test_split(df, test_size=dataset_params["validation_ratio"], stratify=df["speaker_idx"])

Anthony Larcher's avatar
Anthony Larcher committed
969
        torch.manual_seed(dataset_params['seed'])
Anthony Larcher's avatar
Anthony Larcher committed
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
        #training_set = SpkSet(dataset_yaml,
        #                      set_type="train",
        #                      dataset_df=training_df,
        #                      overlap=dataset_params['train']['overlap'],
        #                      output_format=output_format,
        #                      windowed=True)

        training_set = SideSet(dataset_yaml,
                               set_type="train",
                               chunk_per_segment=-1,
                               overlap=dataset_params['train']['overlap'],
                               dataset_df=training_df,
                               output_format=output_format,
                               )

Anthony Larcher's avatar
Anthony Larcher committed
985
986
987
988
        validation_set = SideSet(dataset_yaml,
                                 set_type="validation",
                                 dataset_df=validation_df,
                                 output_format=output_format)
Anthony Larcher's avatar
Anthony Larcher committed
989
990


Anthony Larcher's avatar
Anthony Larcher committed
991
        if write_batches_to_disk:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
992
            logging.critical("Start writing batches on disk")
Anthony Larcher's avatar
Anthony Larcher committed
993
994
            training_set.write_to_disk(dataset_params["batch_size"], train_batch_fn_format, num_thread)
            validation_set.write_to_disk(dataset_params["batch_size"], val_batch_fn_format, num_thread)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
995
            logging.critical("---> Done")
Anthony Larcher's avatar
Anthony Larcher committed
996
997

    if load_batches_from_disk:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
998
999
1000
        training_set = FileSet(train_batch_fn_format)
        validation_set = FileSet(train_batch_fn_format)
        batch_size = 1