xvector.py 75.1 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
Anthony Larcher committed
41
import tqdm
Anthony Larcher's avatar
Anthony Larcher committed
42
43
import yaml

Anthony Larcher's avatar
Anthony Larcher committed
44
from torchvision import transforms
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 .res_net import RawPreprocessor, ResBlockWFMS, ResBlock, PreResNet34
Anthony Larcher's avatar
Anthony Larcher committed
52
from ..bosaris import IdMap
Anthony Larcher's avatar
Anthony Larcher committed
53
54
55
56
from ..bosaris import Key
from ..bosaris import Ndx
from ..bosaris.detplot import rocch
from ..bosaris.detplot import rocch2eer
Anthony Larcher's avatar
Anthony Larcher committed
57
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
58
from ..iv_scoring import cosine_scoring
Anthony Larcher's avatar
Anthony Larcher committed
59
from torch.utils.data import DataLoader
Anthony Larcher's avatar
Anthony Larcher committed
60
from sklearn.model_selection import train_test_split
Anthony Larcher's avatar
Anthony Larcher committed
61
from .sincnet import SincNet
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
62
63
64
from .loss import ArcLinear
from .loss import ArcFace
from .loss import l2_norm
Anthony Larcher's avatar
Anthony Larcher committed
65
from .loss import ArcMarginProduct
Anthony Larcher's avatar
Anthony Larcher committed
66

Anthony Larcher's avatar
ddp    
Anthony Larcher committed
67

Anthony Larcher's avatar
Anthony Larcher committed
68
os.environ['MKL_THREADING_LAYER'] = 'GNU'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
69

Anthony Larcher's avatar
Anthony Larcher committed
70
71
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
v1.3.7    
Anthony Larcher committed
72
__copyright__ = "Copyright 2015-2021 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
73
74
75
76
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
77
78


Anthony Larcher's avatar
Anthony Larcher committed
79
80
logging.basicConfig(format='%(asctime)s %(message)s')

Anthony Larcher's avatar
Anthony Larcher committed
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101

# 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
102
            self.halt(str(value))
Anthony Larcher's avatar
Anthony Larcher committed
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126

    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
127
        plt.imshow(numpy.transpose(npimg, (1, 2, 0)))
Anthony Larcher's avatar
Anthony Larcher committed
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142

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
143

Anthony Larcher's avatar
Anthony Larcher committed
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
    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
164
165
def test_metrics(model,
                 device,
Anthony Larcher's avatar
Anthony Larcher committed
166
167
168
                 speaker_number,
                 num_thread,
                 mixed_precision):
Anthony Larcher's avatar
Anthony Larcher committed
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
    """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
185
186
187
    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
debug    
Anthony Larcher committed
188
    data_root_name='/lium/corpus/base/voxceleb1/test/wav'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
189
190

    transform_pipeline = dict()
Anthony Larcher's avatar
debug    
Anthony Larcher committed
191
192
193
194
195
196
197
198
199
200
    #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
201
202
203
204
205
206

    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
207
208
209
                                 transform_pipeline=transform_pipeline,
                                 num_thread=num_thread,
                                 mixed_precision=mixed_precision)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
210
211
212
213
214
215
216
217
218
219
220
221

    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
222

Anthony Larcher's avatar
Anthony Larcher committed
223

224
def get_lr(optimizer):
Anthony Larcher's avatar
Anthony Larcher committed
225
226
227
228
229
    """

    :param optimizer:
    :return:
    """
230
231
232
233
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
234
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
235
236
237
238
239
240
241
242
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
243
244
245
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
246

Anthony Larcher's avatar
Anthony Larcher committed
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
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
267

Anthony Larcher's avatar
Anthony Larcher committed
268

Anthony Larcher's avatar
Anthony Larcher committed
269
270
271
272
273
274
275
276
277
278
279
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
280
        super(GruPooling, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
        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
303

Anthony Larcher's avatar
Anthony Larcher committed
304
class Xtractor(torch.nn.Module):
305
306
307
    """
    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
308

Anthony Larcher's avatar
Anthony Larcher committed
309
310
311
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
312
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
313
314
315
                 norm_embedding=False,
                 aam_margin=0.5,
                 aam_s=0.5):
Anthony Larcher's avatar
Anthony Larcher committed
316
317
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
318
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
319
        """
Anthony Larcher's avatar
Anthony Larcher committed
320
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
321
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
322
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
323
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
324

Anthony Larcher's avatar
Anthony Larcher committed
325
326
        print(f"Speaker number : {self.speaker_number}")

Anthony Larcher's avatar
Anthony Larcher committed
327
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
328

Anthony Larcher's avatar
Anthony Larcher committed
329
330
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
331
332
333
334
335
            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
336
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
337
338
339
            self.activation = torch.nn.LeakyReLU(0.2)

            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
340

Anthony Larcher's avatar
xv    
Anthony Larcher committed
341
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
342
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
343
                ("activation1", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
344
                ("batch_norm1", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
345
346
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
347
                ("batch_norm2", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
348
349
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
350
                ("batch_norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
351
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
352
                ("activation4", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
353
                ("batch_norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
354
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
355
                ("activation5", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
356
                ("batch_norm5", torch.nn.BatchNorm1d(1536))
Anthony Larcher's avatar
Anthony Larcher committed
357
358
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
359
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
360
            self.stat_pooling_weight_decay = 0
Anthony Larcher's avatar
xv    
Anthony Larcher committed
361
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
362
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
363
364
            ]))

365
366
            self.embedding_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
367
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
368
369
370
371
372
                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
373
374
375
            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
376
                    ("batch_norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
377
378
379
                    ("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
380
                    ("batch_norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
381
382
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
383

Anthony Larcher's avatar
debug    
Anthony Larcher committed
384
385
386
            self.sequence_network_weight_decay = 0.0002
            self.before_speaker_embedding_weight_decay = 0.002
            self.after_speaker_embedding_weight_decay = 0.002
387
            self.embedding_size = 512
Anthony Larcher's avatar
Anthony Larcher committed
388

Anthony Larcher's avatar
Anthony Larcher committed
389
        elif model_archi == "resnet34":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
390
            self.input_nbdim = 2
Anthony Larcher's avatar
Anthony Larcher committed
391
392
393
394
395
396
397
398
399
            self.preprocessor = None
            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

400
            self.embedding_size = 256
Anthony Larcher's avatar
Anthony Larcher committed
401

402
            self.loss = "aam"
Anthony Larcher's avatar
Anthony Larcher committed
403
404
405
            self.after_speaker_embedding = ArcMarginProduct(256,
                                                            int(self.speaker_number),
                                                            s = 30.0,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
406
                                                            m = 0.20,
Anthony Larcher's avatar
Anthony Larcher committed
407
                                                            easy_margin = True)
Anthony Larcher's avatar
Anthony Larcher committed
408
409
410
411
412
413
414
415

            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
416
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
417
418
419
420
421
422

            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
423
424
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
425
426
427
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
428
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
429
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
430
431
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448

            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
449
450
451
452
453
            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
454
            elif self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
455
456
457
                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
458

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
459
460
461
462
463
464
            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
465
        else:
Anthony Larcher's avatar
Anthony Larcher committed
466
467
            is_first_resblock = True

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
468
469
470
471
472
473
            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
474

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
475
            self.loss = cfg["training"]["loss"]
Anthony Larcher's avatar
Anthony Larcher committed
476
            if self.loss == "aam":
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
477
478
                self.aam_margin = cfg["training"]["aam_margin"]
                self.aam_s = cfg["training"]["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
479

Anthony Larcher's avatar
Anthony Larcher committed
480
481
482
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
483
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
484
485
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
486
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
487
488
489
490
491
492
493
494
495
496
497
498
                        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
499
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
500
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
501
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
502
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
503
504
505
506
507
508
                                                        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
509
                self.preprocessor_weight_decay = 0.000
Anthony Larcher's avatar
Anthony Larcher committed
510
511

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
512
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
513
            """
Anthony Larcher's avatar
Anthony Larcher committed
514
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
515
516
517
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
518
519
520
521
522
523
524
525
526
527
528
529
            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()

Anthony Larcher's avatar
Anthony Larcher committed
530
531
532
533
534
            if cfg["segmental"][list(cfg["segmental"].keys())[0]].startswith("conv2D"):
                self.input_nbdim = 3
            elif cfg["segmental"][list(cfg["segmental"].keys())[0]].startswith("conv"):
                self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
535
536
537
            # 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
538
539
540
541
542
                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
543
544
545
546
547
                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
548
                                                                stride=1)))
Anthony Larcher's avatar
Anthony Larcher committed
549

Anthony Larcher's avatar
Anthony Larcher committed
550
                elif k.startswith("conv"):
Anthony Larcher's avatar
Anthony Larcher committed
551
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
552
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
553
554
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
555
556
                    input_size = cfg["segmental"][k]["output_channels"]

Anthony Larcher's avatar
Anthony Larcher committed
557
558
                elif k.startswith("ctrans"):
                    segmental_layers.append((k, torch.nn.ConvTranspose1d(input_size,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
559
                                                                         cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
560
561
                                                                         kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                         dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
562
563
564
                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
565
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
Anthony Larcher committed
566
567
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
568
569
570
571
572
573
                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
574
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
575
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
576

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
577
578
579
580
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
581
            tmp_input_size = input_size * 2
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
582
583
584
585
            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
586
                tmp_input_size = cfg["stat_pooling"]["gru_node"]
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
587

Anthony Larcher's avatar
Anthony Larcher committed
588
589
            self.stat_pooling_weight_decay = cfg["stat_pooling"]["weight_decay"]

Anthony Larcher's avatar
Anthony Larcher committed
590
            """
Anthony Larcher's avatar
Anthony Larcher committed
591
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
592
            """
Anthony Larcher's avatar
Anthony Larcher committed
593
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
594
            input_size = tmp_input_size
Anthony Larcher's avatar
xv    
Anthony Larcher committed
595
596
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
597
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
598
599
                    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
600
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
601
602
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
603
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
604
605

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
606
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
607

Anthony Larcher's avatar
Anthony Larcher committed
608
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
609
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
610
611

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
612
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
613

Anthony Larcher's avatar
Anthony Larcher committed
614
            self.embedding_size = input_size
Anthony Larcher's avatar
Anthony Larcher committed
615
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
616
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
617

Anthony Larcher's avatar
Anthony Larcher committed
618
            # if loss_criteria is "cce"
Anthony Larcher's avatar
xv    
Anthony Larcher committed
619
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
620
621
622
623
624
625
626
627
            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
628
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
629
                            input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
630

Anthony Larcher's avatar
Anthony Larcher committed
631
632
633
634
635
                    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
636

Anthony Larcher's avatar
Anthony Larcher committed
637
638
                    elif k.startswith("activation"):
                        after_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
639

Anthony Larcher's avatar
Anthony Larcher committed
640
641
                    elif k.startswith('batch_norm'):
                        after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
642

Anthony Larcher's avatar
Anthony Larcher committed
643
644
645
646
647
648
                    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
649
650
651
652
653
654
655
                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
656
657
658
659
                #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
660
661
662
663
                #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
664

Anthony Larcher's avatar
Anthony Larcher committed
665
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
666

Anthony Larcher's avatar
Anthony Larcher committed
667

668
    def forward(self, x, is_eval=False, target=None, extract_after_pooling=False):
669
670
671
        """

        :param x:
672
        :param is_eval: False for training
673
674
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
675
676
677
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
678
        x = self.sequence_network(x)
679

Anthony Larcher's avatar
Anthony Larcher committed
680
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
681
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
682

683
684
685
        if extract_after_pooling:
            return x

Anthony Larcher's avatar
Anthony Larcher committed
686
        x = self.before_speaker_embedding(x)
687

Anthony Larcher's avatar
Anthony Larcher committed
688
        if self.norm_embedding:
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
689
            #x_norm = x.norm(p=2,dim=1, keepdim=True) / 10. # Why  10. ?
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
690
691
692
            #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
693

Anthony Larcher's avatar
Anthony Larcher committed
694
        if self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
695
            if is_eval:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
696
                return x
Anthony Larcher's avatar
debug    
Anthony Larcher committed
697
            else:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
698
                return self.after_speaker_embedding(x), x
Anthony Larcher's avatar
Anthony Larcher committed
699

Anthony Larcher's avatar
Anthony Larcher committed
700
        elif self.loss == "aam":
Anthony Larcher's avatar
debug    
Anthony Larcher committed
701
702
            if is_eval:
                x = torch.nn.functional.normalize(x, dim=1)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
703
            else:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
704
                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
705

Anthony Larcher's avatar
Anthony Larcher committed
706
        return x
Anthony Larcher's avatar
Anthony Larcher committed
707

708
709
    def context_size(self):
        context = 1
Anthony Larcher's avatar
Anthony Larcher committed
710
711
712
713
714
715
716
717
718
        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
719

Anthony Larcher's avatar
Anthony Larcher committed
720
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
721
           dataset_yaml,
Anthony Larcher's avatar
Anthony Larcher committed
722
723
           epochs=None,
           lr=None,
Anthony Larcher's avatar
Anthony Larcher committed
724
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
725
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
726
727
728
729
           loss=None,
           aam_margin=None,
           aam_s=None,
           patience=None,
Anthony Larcher's avatar
Anthony Larcher committed
730
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
731
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
732
           multi_gpu=True,
Anthony Larcher's avatar
Anthony Larcher committed
733
           mixed_precision=False,
734
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
735
           opt=None,
Anthony Larcher's avatar
Anthony Larcher committed
736
737
           reset_parts=[],
           freeze_parts=[],
Anthony Larcher's avatar
Anthony Larcher committed
738
739
740
741
           num_thread=None,
           write_batches_to_disk=False,
           load_batches_from_disk=False,
           tmp_batch_dir=None):
742
743
    """

Anthony Larcher's avatar
Anthony Larcher committed
744
745
746
747
748
749
    :param speaker_number:
    :param dataset_yaml:
    :param epochs:
    :param lr:
    :param model_yaml:
    :param model_name:
Anthony Larcher's avatar
Anthony Larcher committed
750
751
752
753
    :param loss:
    :param aam_margin:
    :param aam_s:
    :param patience:
Anthony Larcher's avatar
Anthony Larcher committed
754
755
756
757
    :param tmp_model_name:
    :param best_model_name:
    :param multi_gpu:
    :param clipping:
Anthony Larcher's avatar
Anthony Larcher committed
758
759
760
    :param opt:
    :param reset_parts:
    :param freeze_parts:
Anthony Larcher's avatar
Anthony Larcher committed
761
    :param num_thread:
762
763
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
764
765
766
767
    # 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
768

Anthony Larcher's avatar
Anthony Larcher committed
769
770
771
    if write_batches_to_disk:
        load_batches_from_disk = True

Anthony Larcher's avatar
Anthony Larcher committed
772
    if num_thread is None:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
773
774
        import multiprocessing

Anthony Larcher's avatar
Anthony Larcher committed
775
776
        num_thread = multiprocessing.cpu_count()

Anthony Larcher's avatar
Anthony Larcher committed
777
    logging.critical(f"Use {num_thread} cpus")
Anthony Larcher's avatar
Anthony Larcher committed
778
    logging.critical(f"Start process at {time.strftime('%H:%M:%S', time.localtime())}")
779

Anthony Larcher's avatar
Anthony Larcher committed
780
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
781
    # Start from scratch
782
    if model_name is None and model_yaml in ["xvector", "rawnet2", "resnet34"]:
Anthony Larcher's avatar
Anthony Larcher committed
783
        # Initialize a first model
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
784
        if model_yaml == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
785
            model = Xtractor(speaker_number, "xvector", loss=loss)
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
786
        elif model_yaml == "rawnet2":
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
787
            model = Xtractor(speaker_number, "rawnet2")
788
789
        elif model_yaml == "resnet34":
            model = Xtractor(speaker_number, "resnet34")
Anthony Larcher's avatar
Anthony Larcher committed
790
        model_archi = model_yaml
Anthony Larcher's avatar
Anthony Larcher committed
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
    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
817
        if model_name is None:
Anthony Larcher's avatar
Anthony Larcher committed
818
819
820
            model = Xtractor(speaker_number, model_yaml)

         # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
821
        else:
Anthony Larcher's avatar
Anthony Larcher committed
822
823
824
            # Load the model
            logging.critical(f"*** Load model from = {model_name}")
            checkpoint = torch.load(model_name)
Anthony Larcher's avatar
Anthony Larcher committed
825
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
826

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
827
828
            """
            Here we remove all layers that we don't want to reload
Anthony Larcher's avatar
Anthony Larcher committed
829
        
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
830
831
832
833
834
835
836
837
838
839
840
841
842
            """
            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
843

Anthony Larcher's avatar
Anthony Larcher committed
844
845
846
847
848
849
850
    logging.critical(model)

    logging.critical("model_parameters_count: {:d}".format(
        sum(p.numel()
            for p in model.parameters()
            if p.requires_grad)))

Anthony Larcher's avatar
Anthony Larcher committed
851
    embedding_size = model.embedding_size
Anthony Larcher's avatar
Anthony Larcher committed
852

Anthony Larcher's avatar
Anthony Larcher committed
853
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
854
855
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
856
        #model = DDP(model)
Anthony Larcher's avatar
Anthony Larcher committed
857
858
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
859
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
860

Anthony Larcher's avatar
debug    
Anthony Larcher committed
861
862
863
864
    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
865
866
867
    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"
868

Anthony Larcher's avatar
Anthony Larcher committed
869
870
871
872
873
874
875
    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
876

Anthony Larcher's avatar
minor    
Anthony Larcher committed
877
        if write_batches_to_disk or dataset_params["batch_size"] > 1:
Anthony Larcher's avatar
Anthony Larcher committed
878
879
880
881
            output_format = "numpy"
        else:
            output_format = "pytorch"

Anthony Larcher's avatar
Anthony Larcher committed
882
883
        training_df, validation_df = train_test_split(df, test_size=dataset_params["validation_ratio"])
        torch.manual_seed(dataset_params['seed'])
Anthony Larcher's avatar
spkset    
Anthony Larcher committed
884
885
886
887
888
889
        training_set = SpkSet(dataset_yaml,
                              set_type="train",
                              dataset_df=training_df,
                              overlap=dataset_params['train']['overlap'],
                              output_format="pytorch",
                              windowed=True)
Anthony Larcher's avatar
Anthony Larcher committed
890

Anthony Larcher's avatar
Anthony Larcher committed
891
892
893
894
        validation_set = SideSet(dataset_yaml,
                                 set_type="validation",
                                 dataset_df=validation_df,
                                 output_format=output_format)
Anthony Larcher's avatar
Anthony Larcher committed
895
896


Anthony Larcher's avatar
Anthony Larcher committed
897
        if write_batches_to_disk:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
898
            logging.critical("Start writing batches on disk")
Anthony Larcher's avatar
Anthony Larcher committed
899
900
            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
901
            logging.critical("---> Done")
Anthony Larcher's avatar
Anthony Larcher committed
902
903

    if load_batches_from_disk:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
904
905
906
907
908
909
910
911
912
913
914
915
916
        training_set = FileSet(train_batch_fn_format)
        validation_set = FileSet(train_batch_fn_format)
        batch_size = 1
    else:
        batch_size = dataset_params["batch_size"]


    print(f"Size of batches = {batch_size}")
    training_loader = DataLoader(training_set,
                                 batch_size=batch_size,
                                 shuffle=True,
                                 drop_last=True,
                                 pin_memory=True,
Anthony Larcher's avatar
spkset    
Anthony Larcher committed
917
                                 num_workers=num_thread,
Anthony Larcher's avatar
Anthony Larcher committed
918
                                 persistent_workers=True)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
919
920
921

    validation_loader = DataLoader(validation_set,
                                   batch_size=batch_size,
Anthony Larcher's avatar
Anthony Larcher committed
922
                                   drop_last=False,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
923
                                   pin_memory=True,
Anthony Larcher's avatar
spkset    
Anthony Larcher committed
924
                                   num_workers=num_thread,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
925
                                   persistent_workers=False)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
926

Anthony Larcher's avatar
Anthony Larcher committed
927
928
929
    """
    Set the training options
    """
Anthony Larcher's avatar
Anthony Larcher committed
930
    if opt == 'adam':
Anthony Larcher's avatar
Anthony Larcher committed
931
        _optimizer = torch.optim.Adam
Anthony Larcher's avatar
Anthony Larcher committed
932
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
933
934
    elif opt == 'rmsprop':
        _optimizer = torch.optim.RMSprop
Anthony Larcher's avatar
Anthony Larcher committed
935
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
936
937
938
    else: # opt == 'sgd'
        _optimizer = torch.optim.SGD
        _options = {'lr': lr, 'momentum': 0.9}
Anthony Larcher's avatar
Anthony Larcher committed
939

Anthony Larcher's avatar
Anthony Larcher committed
940
    param_list = []
Anthony Larcher's avatar
Anthony Larcher committed
941
    if type(model) is Xtractor:
Anthony Larcher's avatar
Anthony Larcher committed
942
943
944
945
946
947
948
        if model.preprocessor is not None:
            param_list.append({'params': model.preprocessor.parameters(), 'weight_decay': model.preprocessor_weight_decay})
        param_list.append({'params': model.sequence_network.parameters(), 'weight_decay': model.sequence_network_weight_decay})
        param_list.append({'params': model.stat_pooling.parameters(), 'weight_decay': model.stat_pooling_weight_decay})
        param_list.append({'params': model.before_speaker_embedding.parameters(), 'weight_decay': model.before_speaker_embedding_weight_decay})
        param_list.append({'params': model.after_speaker_embedding.parameters(), 'weight_decay': model.after_speaker_embedding_weight_decay})

Anthony Larcher's avatar
Anthony Larcher committed
949
    else:
Anthony Larcher's avatar
Anthony Larcher committed
950
951
952
953
954
955
956
957
        if model.module.preprocessor is not None:
            param_list.append({'params': model.module.preprocessor.parameters(), 'weight_decay': model.module.preprocessor_weight_decay})
        param_list.append({'params': model.module.sequence_network.parameters(), 'weight_decay': model.module.sequence_network_weight_decay})
        param_list.append({'params': model.module.stat_pooling.parameters(), 'weight_decay': model.module.stat_pooling_weight_decay})
        param_list.append({'params': model.module.before_speaker_embedding.parameters(), 'weight_decay': model.module.before_speaker_embedding_weight_decay})
        param_list.append({'params': model.module.after_speaker_embedding.parameters(), 'weight_decay': model.module.after_speaker_embedding_weight_decay})

    optimizer = _optimizer(param_list, **_options)
Anthony Larcher's avatar
Anthony Larcher committed
958
    scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
Anthony Larcher's avatar
Anthony Larcher committed
959
                                                     milestones=numpy.arange(50000,160000,10000),
Anthony Larcher's avatar
Anthony Larcher committed
960
961
                                                     gamma=0.1,
                                                     last_epoch=-1,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
962
                                                     verbose=False)
963

Anthony Larcher's avatar
Anthony Larcher committed
964
965
966
967
968
    if mixed_precision:
        scaler = torch.cuda.amp.GradScaler()
    else:
        scaler = None

Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
969
    best_accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
970
    best_accuracy_epoch = 1
Anthony Larcher's avatar
spkset    
Anthony Larcher committed
971
    best_eer = 100
Anthony Larcher's avatar
Anthony Larcher committed
972
    curr_patience = patience
Anthony Larcher's avatar
eer    
Anthony Larcher committed
973
974
    
    logging.critical("Compute EER before starting")
Anthony Larcher's avatar
debug    
Anthony Larcher committed
975
976
977
978
    val_acc, val_loss, val_eer = cross_validation(model,
                                                  validation_loader,
                                                  device,
                                                  [validation_set.__len__(),
Anthony Larcher's avatar
Anthony Larcher committed
979
980
                                                   embedding_size],
                                                   mixed_precision)
Anthony Larcher's avatar
Anthony Larcher committed
981

Anthony Larcher's avatar
Anthony Larcher committed
982
    test_eer = test_metrics(model, device, speaker_number, num_thread, mixed_precision)
Anthony Larcher's avatar
Anthony Larcher committed
983

Anthony Larcher's avatar
debug    
Anthony Larcher committed
984
985
    logging.critical(f"***{time.strftime('%H:%M:%S', time.localtime())} Initial metrics - Cross validation accuracy = {val_acc} %, EER = {val_eer * 100} %")
    logging.critical(f"***{time.strftime('%H:%M:%S', time.localtime())} Initial metrics - Test EER = {test_eer * 100} %")
Anthony Larcher's avatar
Anthony Larcher committed
986

Anthony Larcher's avatar
Anthony Larcher committed
987
    for epoch in range(1, epochs + 1):
988
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
989
990
991
        if curr_patience == 0:
            print(f"Stopping at epoch {epoch} for cause of patience")
            break
Anthony Larcher's avatar
Anthony Larcher committed
992
993
994
995
        model = train_epoch(model,
                            epoch,
                            training_loader,
                            optimizer,
Anthony Larcher's avatar
Anthony Larcher committed
996
                            scheduler,
Anthony Larcher's avatar
Anthony Larcher committed
997
                            dataset_params["log_interval"],
Anthony Larcher's avatar
Anthony Larcher committed
998
999
                            device,
                            scaler=scaler,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
1000
                            clipping=clipping)