xvector.py 81.2 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
87
88
89
90
91
92
93
94
95
96
97
98
99

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


class GuruMeditation (torch.autograd.detect_anomaly):
    
    def __init__(self):
        super(GuruMeditation, self).__init__()

    def __enter__(self):
        super(GuruMeditation, self).__enter__()
        return self

    def __exit__(self, type, value, trace):
        super(GuruMeditation, self).__exit__()
        if isinstance(value, RuntimeError):
            traceback.print_tb(trace)
Anthony Larcher's avatar
Anthony Larcher committed
100
            self.halt(str(value))
Anthony Larcher's avatar
Anthony Larcher committed
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124

    def halt(msg):
        print (msg)
        pdb.set_trace()


def select_n_random(data, labels, n=100):
    '''
    Selects n random datapoints and their corresponding labels from a dataset
    '''
    assert len(data) == len(labels)

    perm = torch.randperm(len(data))
    return data[perm][:n], labels[perm][:n]


def matplotlib_imshow(img, one_channel=False):
    if one_channel:
        img = img.mean(dim=0)
    img = img / 2 + 0.5     # unnormalize
    npimg = img.cpu().numpy()
    if one_channel:
        plt.imshow(npimg, cmap="Greys")
    else:
Anthony Larcher's avatar
Anthony Larcher committed
125
        plt.imshow(numpy.transpose(npimg, (1, 2, 0)))
Anthony Larcher's avatar
Anthony Larcher committed
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140

def speech_to_probs(model, speech):
    '''
    Generates predictions and corresponding probabilities from a trained
    network and a list of images
    '''
    output = model(speech)
    # convert output probabilities to predicted class
    _, preds_tensor = torch.max(output, 1)
    preds = numpy.squeeze(preds_tensor.cpu().numpy())
    return preds, [torch.nn.functional.softmax(el, dim=0)[i].item() for i, el in zip(preds, output)]


def plot_classes_preds(model, speech, labels):
    '''
Anthony Larcher's avatar
Anthony Larcher committed
141

Anthony Larcher's avatar
Anthony Larcher committed
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
    Generates matplotlib Figure using a trained network, along with images
    and labels from a batch, that shows the network's top prediction along
    with its probability, alongside the actual label, coloring this
    information based on whether the prediction was correct or not.
    Uses the "speech_to_probs" function.
    '''
    preds, probs = speech_to_probs(model, speech)
    # plot the images in the batch, along with predicted and true labels
    fig = plt.figure(figsize=(12, 48))
    for idx in numpy.arange(4):
         ax = fig.add_subplot(1, 4, idx+1, xticks=[], yticks=[])
         #matplotlib_imshow(speech[idx], one_channel=True)
         ax.set_title("{0}, {1:.1f}%\n(label: {2})".format(
                                                           preds[idx],
                                                           probs[idx] * 100.0,
                                                           labels[idx]),
                                                           color=("green" if preds[idx]==labels[idx].item() else "red"))
    return fig


Anthony Larcher's avatar
debug    
Anthony Larcher committed
162
163
def test_metrics(model,
                 device,
Anthony Larcher's avatar
Anthony Larcher committed
164
165
166
                 speaker_number,
                 num_thread,
                 mixed_precision):
Anthony Larcher's avatar
Anthony Larcher committed
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
    """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
183
184
185
    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
186
    data_root_name='/data/larcher/voxceleb1/test/wav'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
187
188

    transform_pipeline = dict()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
189
190
191
192
193
194
195
196
197
198
    #mfcc_config = dict()
    #mfcc_config['nb_filters'] = 81
    #mfcc_config['nb_ceps'] = 80
    #mfcc_config['lowfreq'] = 133.333
    #mfcc_config['maxfreq'] = 6855.4976
    #mfcc_config['win_time'] = 0.025
    #mfcc_config['shift'] = 0.01
    #mfcc_config['n_fft'] = 2048
    #transform_pipeline['MFCC'] = mfcc_config
    #transform_pipeline['CMVN'] = {}
Anthony Larcher's avatar
debug    
Anthony Larcher committed
199
200
201
202
203
204

    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
205
206
207
                                 transform_pipeline=transform_pipeline,
                                 num_thread=num_thread,
                                 mixed_precision=mixed_precision)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
208
209
210
211
212
213
214
215
216
217
218
219

    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
220

Anthony Larcher's avatar
Anthony Larcher committed
221

222
def get_lr(optimizer):
Anthony Larcher's avatar
Anthony Larcher committed
223
224
225
226
227
    """

    :param optimizer:
    :return:
    """
228
229
230
231
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
232
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
233
234
235
236
237
238
239
240
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
241
242
243
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
244

Anthony Larcher's avatar
Anthony Larcher committed
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
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
265

Anthony Larcher's avatar
Anthony Larcher committed
266

Anthony Larcher's avatar
Anthony Larcher committed
267
268
269
270
271
272
273
274
275
276
277
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
278
        super(GruPooling, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
        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
301

Anthony Larcher's avatar
Anthony Larcher committed
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
class PreEmphasis(torch.nn.Module):

    def __init__(self, coef: float = 0.97):
        super().__init__()
        self.coef = coef
        # make kernel
        # In pytorch, the convolution operation uses cross-correlation. So, filter is flipped.
        self.register_buffer(
            'flipped_filter', torch.FloatTensor([-self.coef, 1.]).unsqueeze(0).unsqueeze(0)
        )

    def forward(self, input: torch.tensor) -> torch.tensor:
        assert len(input.size()) == 2, 'The number of dimensions of input tensor must be 2!'
        # reflect padding to match lengths of in/out
        input = input.unsqueeze(1)
        input = torch.nn.functional.pad(input, (1, 0), 'reflect')
        return torch.nn.functional.conv1d(input, self.flipped_filter).squeeze(1)


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

    """

    def __init__(self,
                 pre_emphasis=0.97,
                 sample_rate=16000,
Anthony Larcher's avatar
Anthony Larcher committed
329
                 n_fft=2048,
Anthony Larcher's avatar
Anthony Larcher committed
330
331
332
333
                 f_min=133.333,
                 f_max=6855.4976,
                 win_length=1024,
                 window_fn=torch.hann_window,
Anthony Larcher's avatar
Anthony Larcher committed
334
                 hop_length=512,
Anthony Larcher's avatar
Anthony Larcher committed
335
                 power=2.0,
Anthony Larcher's avatar
Anthony Larcher committed
336
                 n_mels=100,
Anthony Larcher's avatar
Anthony Larcher committed
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
                 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
356
357
358
359
360
361
362
363
        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
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386

        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
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456

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
457
class Xtractor(torch.nn.Module):
458
459
460
    """
    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
461

Anthony Larcher's avatar
Anthony Larcher committed
462
463
464
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
465
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
466
                 norm_embedding=False,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
467
468
                 aam_margin=0.2,
                 aam_s=30):
Anthony Larcher's avatar
Anthony Larcher committed
469
470
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
471
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
472
        """
Anthony Larcher's avatar
Anthony Larcher committed
473
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
474
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
475
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
476
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
477

Anthony Larcher's avatar
Anthony Larcher committed
478
479
        print(f"Speaker number : {self.speaker_number}")

Anthony Larcher's avatar
Anthony Larcher committed
480
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
481

Anthony Larcher's avatar
Anthony Larcher committed
482
483
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
484
485
486
487
488
            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
489
490
            self.activation = torch.nn.LeakyReLU(0.2)

Anthony Larcher's avatar
Anthony Larcher committed
491
            self.preprocessor = MfccFrontEnd()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
492
            self.feature_size = self.preprocessor.n_mfcc
Anthony Larcher's avatar
Anthony Larcher committed
493

Anthony Larcher's avatar
xv    
Anthony Larcher committed
494
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
495
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
496
                ("activation1", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
497
                ("batch_norm1", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
498
499
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
500
                ("batch_norm2", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
501
502
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
503
                ("batch_norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
504
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
505
                ("activation4", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
506
                ("batch_norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
507
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
508
                ("activation5", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
509
                ("batch_norm5", torch.nn.BatchNorm1d(1536))
Anthony Larcher's avatar
Anthony Larcher committed
510
511
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
512
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
513
            self.stat_pooling_weight_decay = 0
Anthony Larcher's avatar
xv    
Anthony Larcher committed
514
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
515
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
516
517
            ]))

518
519
            self.embedding_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
520
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
521
522
523
524
525
                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
526
527
528
            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
529
                    ("batch_norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
530
531
532
                    ("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
533
                    ("batch_norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
534
535
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
536

Anthony Larcher's avatar
debug    
Anthony Larcher committed
537
538
539
            self.sequence_network_weight_decay = 0.0002
            self.before_speaker_embedding_weight_decay = 0.002
            self.after_speaker_embedding_weight_decay = 0.002
540
            self.embedding_size = 512
Anthony Larcher's avatar
Anthony Larcher committed
541

Anthony Larcher's avatar
Anthony Larcher committed
542
        elif model_archi == "resnet34":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
543

Anthony Larcher's avatar
Anthony Larcher committed
544
            self.preprocessor = MfccFrontEnd()
Anthony Larcher's avatar
Anthony Larcher committed
545
546
547
548
549
550
551
552
            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

553
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
554

555
            self.loss = "aam"
Anthony Larcher's avatar
Anthony Larcher committed
556
557
558
            self.after_speaker_embedding = ArcMarginProduct(256,
                                                            int(self.speaker_number),
                                                            s = 30.0,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
559
                                                            m = 0.20,
Anthony Larcher's avatar
Anthony Larcher committed
560
                                                            easy_margin = True)
Anthony Larcher's avatar
Anthony Larcher committed
561
562
563
564
565
566
567

            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
568
        elif model_archi == "fastresnet34":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
569

Anthony Larcher's avatar
FB    
Anthony Larcher committed
570
            self.preprocessor = MelSpecFrontEnd()
Anthony Larcher's avatar
Anthony Larcher committed
571
            self.sequence_network = PreFastResNet34()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
572
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
573

Anthony Larcher's avatar
Anthony Larcher committed
574
            self.before_speaker_embedding = torch.nn.Linear(in_features = 2560,
Anthony Larcher's avatar
Anthony Larcher committed
575
                                                            out_features = self.embedding_size)
Anthony Larcher's avatar
Anthony Larcher committed
576
577
578
579
580

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

            self.loss = "aam"
Anthony Larcher's avatar
debug    
Anthony Larcher committed
581
            self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
Anthony Larcher's avatar
Anthony Larcher committed
582
                                                            int(self.speaker_number),
Anthony Larcher's avatar
debug    
Anthony Larcher committed
583
584
                                                            s = 30,
                                                            m = 0.2,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
585
                                                            easy_margin = False)
Anthony Larcher's avatar
Anthony Larcher committed
586
587
588
589
590
591

            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
592

Anthony Larcher's avatar
Anthony Larcher committed
593
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
594
595
596
597
598
599

            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
600
601
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
602
603
604
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
605
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
606
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
607
608
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625

            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
626
627
628
629
630
            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
631
            elif self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
632
633
634
                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
635

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
636
637
638
639
640
641
            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
642
        else:
Anthony Larcher's avatar
Anthony Larcher committed
643
644
            is_first_resblock = True

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
645
646
647
648
649
650
            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
651

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
652
            self.loss = cfg["training"]["loss"]
Anthony Larcher's avatar
Anthony Larcher committed
653
            if self.loss == "aam":
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
654
655
                self.aam_margin = cfg["training"]["aam_margin"]
                self.aam_s = cfg["training"]["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
656

Anthony Larcher's avatar
Anthony Larcher committed
657
658
659
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
660
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
661
662
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
663
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
664
665
666
667
668
669
670
671
672
673
674
675
                        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
676
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
677
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
678
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
679
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
680
681
682
683
684
685
                                                        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
686
                self.preprocessor_weight_decay = 0.000
Anthony Larcher's avatar
Anthony Larcher committed
687
688

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
689
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
690
            """
Anthony Larcher's avatar
Anthony Larcher committed
691
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
692
693
694
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
            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
710
711
712
713
714
                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
715
716
717
718
719
                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
720
                                                                stride=1)))
Anthony Larcher's avatar
Anthony Larcher committed
721

Anthony Larcher's avatar
Anthony Larcher committed
722
                elif k.startswith("conv"):
Anthony Larcher's avatar
Anthony Larcher committed
723
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
724
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
725
726
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
727
728
                    input_size = cfg["segmental"][k]["output_channels"]

Anthony Larcher's avatar
Anthony Larcher committed
729
730
                elif k.startswith("ctrans"):
                    segmental_layers.append((k, torch.nn.ConvTranspose1d(input_size,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
731
                                                                         cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
732
733
                                                                         kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                         dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
734
735
736
                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
737
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
Anthony Larcher committed
738
739
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
740
741
742
743
744
745
                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
746
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
747
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
748

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
749
750
751
752
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
753
            tmp_input_size = input_size * 2
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
754
755
756
757
            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
758
                tmp_input_size = cfg["stat_pooling"]["gru_node"]
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
759

Anthony Larcher's avatar
Anthony Larcher committed
760
761
            self.stat_pooling_weight_decay = cfg["stat_pooling"]["weight_decay"]

Anthony Larcher's avatar
Anthony Larcher committed
762
            """
Anthony Larcher's avatar
Anthony Larcher committed
763
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
764
            """
Anthony Larcher's avatar
Anthony Larcher committed
765
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
766
            input_size = tmp_input_size
Anthony Larcher's avatar
xv    
Anthony Larcher committed
767
768
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
769
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
770
771
                    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
772
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
773
774
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
775
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
776
777

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
778
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
779

Anthony Larcher's avatar
Anthony Larcher committed
780
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
781
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
782
783

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
784
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
785

Anthony Larcher's avatar
Anthony Larcher committed
786
            self.embedding_size = input_size
Anthony Larcher's avatar
Anthony Larcher committed
787
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
788
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
789

Anthony Larcher's avatar
Anthony Larcher committed
790
            # if loss_criteria is "cce"
Anthony Larcher's avatar
xv    
Anthony Larcher committed
791
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
792
793
794
795
796
797
798
799
            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
800
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
801
                            input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
802

Anthony Larcher's avatar
Anthony Larcher committed
803
804
805
806
807
                    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
808

Anthony Larcher's avatar
Anthony Larcher committed
809
810
                    elif k.startswith("activation"):
                        after_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
811

Anthony Larcher's avatar
Anthony Larcher committed
812
813
                    elif k.startswith('batch_norm'):
                        after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
814

Anthony Larcher's avatar
Anthony Larcher committed
815
816
817
818
819
820
                    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
821
822
823
824
825
826
827
                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
828
829
830
831
                #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
832
833
834
835
                #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
836

Anthony Larcher's avatar
Anthony Larcher committed
837
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
838

Anthony Larcher's avatar
Anthony Larcher committed
839

840
    def forward(self, x, is_eval=False, target=None, extract_after_pooling=False):
841
842
843
        """

        :param x:
844
        :param is_eval: False for training
845
846
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
847
848
849
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
850
        x = self.sequence_network(x)
851

Anthony Larcher's avatar
Anthony Larcher committed
852
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
853
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
854

855
856
857
        if extract_after_pooling:
            return x

Anthony Larcher's avatar
Anthony Larcher committed
858
        x = self.before_speaker_embedding(x)
859

Anthony Larcher's avatar
Anthony Larcher committed
860
        if self.norm_embedding:
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
861
            #x_norm = x.norm(p=2,dim=1, keepdim=True) / 10. # Why  10. ?
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
862
863
864
            #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
865

Anthony Larcher's avatar
Anthony Larcher committed
866
        if self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
867
            if is_eval:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
868
                return x
Anthony Larcher's avatar
debug    
Anthony Larcher committed
869
            else:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
870
                return self.after_speaker_embedding(x), x
Anthony Larcher's avatar
Anthony Larcher committed
871

Anthony Larcher's avatar
Anthony Larcher committed
872
        elif self.loss == "aam":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
873
874
            if is_eval:
                x = torch.nn.functional.normalize(x, dim=1)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
875
            else:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
876
                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
877

Anthony Larcher's avatar
Anthony Larcher committed
878
        return x
Anthony Larcher's avatar
Anthony Larcher committed
879

880
881
    def context_size(self):
        context = 1
Anthony Larcher's avatar
Anthony Larcher committed
882
883
884
885
886
887
888
889
890
        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
891

Anthony Larcher's avatar
Anthony Larcher committed
892
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
893
           dataset_yaml,
Anthony Larcher's avatar
Anthony Larcher committed
894
895
           epochs=None,
           lr=None,
Anthony Larcher's avatar
Anthony Larcher committed
896
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
897
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
898
899
900
901
           loss=None,
           aam_margin=None,
           aam_s=None,
           patience=None,
Anthony Larcher's avatar
Anthony Larcher committed
902
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
903
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
904
           multi_gpu=True,
Anthony Larcher's avatar
Anthony Larcher committed
905
           mixed_precision=False,
906
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
907
           opt=None,
Anthony Larcher's avatar
Anthony Larcher committed
908
909
           reset_parts=[],
           freeze_parts=[],
Anthony Larcher's avatar
Anthony Larcher committed
910
911
912
           num_thread=None,
           write_batches_to_disk=False,
           load_batches_from_disk=False,
Anthony Larcher's avatar
Anthony Larcher committed
913
914
           tmp_batch_dir=None,
           compute_test_eer=True):
915
916
    """

Anthony Larcher's avatar
Anthony Larcher committed
917
918
919
920
921
922
    :param speaker_number:
    :param dataset_yaml:
    :param epochs:
    :param lr:
    :param model_yaml:
    :param model_name:
Anthony Larcher's avatar
Anthony Larcher committed
923
924
925
926
    :param loss:
    :param aam_margin:
    :param aam_s:
    :param patience:
Anthony Larcher's avatar
Anthony Larcher committed
927
928
929
930
    :param tmp_model_name:
    :param best_model_name:
    :param multi_gpu:
    :param clipping:
Anthony Larcher's avatar
Anthony Larcher committed
931
932
933
    :param opt:
    :param reset_parts:
    :param freeze_parts:
Anthony Larcher's avatar
Anthony Larcher committed
934
    :param num_thread:
935
936
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
937
938
939
940
    # 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
941

Anthony Larcher's avatar
debug    
Anthony Larcher committed
942
943
944
    # Test to optimize
    torch.autograd.profiler.emit_nvtx(enabled=False)

Anthony Larcher's avatar
Anthony Larcher committed
945
946
947
    if write_batches_to_disk:
        load_batches_from_disk = True

Anthony Larcher's avatar
Anthony Larcher committed
948
    if num_thread is None:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
949
        import multiprocessing
Anthony Larcher's avatar
Anthony Larcher committed
950
951
        num_thread = multiprocessing.cpu_count()

Anthony Larcher's avatar
Anthony Larcher committed
952
    logging.critical(f"Use {num_thread} cpus")
Anthony Larcher's avatar
Anthony Larcher committed
953
    logging.critical(f"Start process at {time.strftime('%H:%M:%S', time.localtime())}")
954

Anthony Larcher's avatar
debug    
Anthony Larcher committed
955
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
956
    # Start from scratch
Anthony Larcher's avatar
Anthony Larcher committed
957
    if model_name is None and model_yaml in ["xvector", "rawnet2", "resnet34", "fastresnet34"]:
Anthony Larcher's avatar
Anthony Larcher committed
958
        # Initialize a first model
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
959
        if model_yaml == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
960
            model = Xtractor(speaker_number, "xvector", loss=loss)
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
961
        elif model_yaml == "rawnet2":
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
962
            model = Xtractor(speaker_number, "rawnet2")
963
964
        elif model_yaml == "resnet34":
            model = Xtractor(speaker_number, "resnet34")
Anthony Larcher's avatar
Anthony Larcher committed
965
966
        elif model_yaml == "fastresnet34":
            model = Xtractor(speaker_number, "fastresnet34")
Anthony Larcher's avatar
Anthony Larcher committed
967
        model_archi = model_yaml
Anthony Larcher's avatar
Anthony Larcher committed
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
    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
994
        if model_name is None:
Anthony Larcher's avatar
Anthony Larcher committed
995
996
997
            model = Xtractor(speaker_number, model_yaml)

         # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
998
        else:
Anthony Larcher's avatar
Anthony Larcher committed
999
1000
            # Load the model
            logging.critical(f"*** Load model from = {model_name}")
For faster browsing, not all history is shown. View entire blame