xvector.py 24.6 KB
Newer Older
Anthony Larcher's avatar
Anthony Larcher committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
# -*- coding: utf-8 -*-
#
# 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
Anthony Larcher committed
25
Copyright 2014-2020 Yevhenii Prokopalo, Anthony Larcher
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
import logging
import numpy
Anthony Larcher's avatar
Anthony Larcher committed
30
import pandas
Anthony Larcher's avatar
minor    
Anthony Larcher committed
31
import pickle
Anthony Larcher's avatar
Anthony Larcher committed
32
import shutil
Anthony Larcher's avatar
Anthony Larcher committed
33
import torch
Anthony Larcher's avatar
Anthony Larcher committed
34
35
import torch.optim as optim
import torch.multiprocessing as mp
Anthony Larcher's avatar
Anthony Larcher committed
36
37
import yaml

Anthony Larcher's avatar
Anthony Larcher committed
38
from torchvision import transforms
Anthony Larcher's avatar
Anthony Larcher committed
39
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
40
from .xsets import XvectorMultiDataset, StatDataset, VoxDataset, SideSet
Anthony Larcher's avatar
Anthony Larcher committed
41
from .xsets import IdMapSet
Anthony Larcher's avatar
Anthony Larcher committed
42
from .xsets import FrequencyMask, CMVN, TemporalMask, MFCC
Anthony Larcher's avatar
Anthony Larcher committed
43
44
from ..bosaris import IdMap
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
45
from torch.utils.data import DataLoader
Anthony Larcher's avatar
Anthony Larcher committed
46
from sklearn.model_selection import train_test_split
Anthony Larcher's avatar
Anthony Larcher committed
47
from .sincnet import SincNet
Anthony Larcher's avatar
Anthony Larcher committed
48

Anthony Larcher's avatar
Anthony Larcher committed
49
50
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
51
__copyright__ = "Copyright 2015-2020 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
52
53
54
55
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
56
57


58
def get_lr(optimizer):
Anthony Larcher's avatar
Anthony Larcher committed
59
60
61
62
63
    """

    :param optimizer:
    :return:
    """
64
65
66
67
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
68
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
69
70
71
72
73
74
75
76
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
77
78
79
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
80
81
82


class Xtractor(torch.nn.Module):
83
84
85
    """
    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
86

Anthony Larcher's avatar
Anthony Larcher committed
87
    def __init__(self, speaker_number, model_archi=None):
Anthony Larcher's avatar
Anthony Larcher committed
88
89
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
90
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
91
        """
Anthony Larcher's avatar
Anthony Larcher committed
92
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
93
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
94
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
95

Anthony Larcher's avatar
Anthony Larcher committed
96
        if model_archi is None:
Anthony Larcher's avatar
Anthony Larcher committed
97
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
98
99
100
            self.activation = torch.nn.LeakyReLU(0.2)

            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
101

Anthony Larcher's avatar
xv    
Anthony Larcher committed
102
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
103
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
104
105
106
107
108
109
110
111
                ("activation1", torch.nn.LeakyReLU(0.2)),
                ("norm1", torch.nn.BatchNorm1d(512)),
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
                ("norm2", torch.nn.BatchNorm1d(512)),
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
                ("norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
112
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
113
114
                ("activation4", torch.nn.LeakyReLU(0.2)),
                ("norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
115
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
116
117
118
119
                ("activation5", torch.nn.LeakyReLU(0.2)),
                ("norm5", torch.nn.BatchNorm1d(1536))
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
120
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
121
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
122
123
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
124
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
125
126
                ("activation6", torch.nn.LeakyReLU(0.2)),
                ("norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
127
                ("dropout6", torch.nn.Dropout(p=0.05)),
Anthony Larcher's avatar
Anthony Larcher committed
128
                ("linear7", torch.nn.Linear(512, 512)),
Anthony Larcher's avatar
Anthony Larcher committed
129
130
                ("activation7", torch.nn.LeakyReLU(0.2)),
                ("norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
131
                ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
Anthony Larcher's avatar
Anthony Larcher committed
132
133
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
134
135
136
137
            self.sequence_network_weight_decay = 0.0002
            self.before_speaker_embedding_weight_decay = 0.002
            self.after_speaker_embedding_weight_decay = 0.002

Anthony Larcher's avatar
Anthony Larcher committed
138
139
        else:
            # Load Yaml configuration
Anthony Larcher's avatar
Anthony Larcher committed
140
            with open(model_archi, 'r') as fh:
Anthony Larcher's avatar
Anthony Larcher committed
141
142
                cfg = yaml.load(fh, Loader=yaml.FullLoader)

Anthony Larcher's avatar
Anthony Larcher committed
143
144
145
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
146
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
147
148
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
149
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
150
151
152
153
154
155
156
157
158
159
160
161
                        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
162
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
163
164

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
165
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
166
            """
Anthony Larcher's avatar
Anthony Larcher committed
167
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
168
169
170
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
            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():
                if k.startswith("conv"):
Anthony Larcher's avatar
Anthony Larcher committed
187
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
188
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
189
190
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
191
192
193
194
195
196
197
198
                    input_size = cfg["segmental"][k]["output_channels"]

                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

                elif k.startswith('norm'):
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
199
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
200
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
201

Anthony Larcher's avatar
Anthony Larcher committed
202
203
204
            """
            Prepapre last part of the network (after pooling)
            """
Anthony Larcher's avatar
Anthony Larcher committed
205
206
            # Create sequential object for the second part of the network
            input_size = input_size * 2
Anthony Larcher's avatar
xv    
Anthony Larcher committed
207
208
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
209
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
210
211
                    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
212
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
213
214
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
215
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
216
217

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
218
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
219
220

                elif k.startswith('norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
221
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
222
223

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
224
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
225

Anthony Larcher's avatar
Anthony Larcher committed
226
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
227
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
228
229
230
231
232

            # Create sequential object for the second part of the network
            after_embedding_layers = []
            for k in cfg["after_embedding"].keys():
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
233
234
                    if cfg["after_embedding"][k]["output"] == "speaker_number":
                        after_embedding_layers.append((k, torch.nn.Linear(input_size, self.speaker_number)))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
235
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
236
237
                        after_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
238
                        input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
239
240
241
242
243
244
245
246

                elif k.startswith("activation"):
                    after_embedding_layers.append((k, self.activation))

                elif k.startswith('norm'):
                    after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
247
                    after_embedding_layers.append((k, torch.nn.Dropout(p=cfg["after_embedding"][k])))
Anthony Larcher's avatar
Anthony Larcher committed
248

Anthony Larcher's avatar
Anthony Larcher committed
249
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict(after_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
250
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
251

Anthony Larcher's avatar
Anthony Larcher committed
252
    def forward(self, x, is_eval=False):
253
254
255
        """

        :param x:
Anthony Larcher's avatar
Anthony Larcher committed
256
        :param is_eval:
257
258
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
259
260
261
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
262
        x = self.sequence_network(x)
263

Anthony Larcher's avatar
Anthony Larcher committed
264
265
266
267
268
269
270
271
        # Mean and Standard deviation pooling
        mean = torch.mean(x, dim=2)
        std = torch.std(x, dim=2)
        x = torch.cat([mean, std], dim=1)

        x = self.before_speaker_embedding(x)
        if is_eval:
            return x
272

Anthony Larcher's avatar
Anthony Larcher committed
273
274
        x = self.after_speaker_embedding(x)
        return x
Anthony Larcher's avatar
Anthony Larcher committed
275

Anthony Larcher's avatar
Anthony Larcher committed
276

Anthony Larcher's avatar
Anthony Larcher committed
277
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
278
           dataset_yaml,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
279
           epochs=100,
Anthony Larcher's avatar
Anthony Larcher committed
280
           lr=0.01,
Anthony Larcher's avatar
Anthony Larcher committed
281
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
282
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
283
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
284
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
285
           multi_gpu=True,
286
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
287
           num_thread=1):
288
289
    """

Anthony Larcher's avatar
Anthony Larcher committed
290
291
292
293
294
295
296
297
298
299
300
    :param speaker_number:
    :param dataset_yaml:
    :param epochs:
    :param lr:
    :param model_yaml:
    :param model_name:
    :param tmp_model_name:
    :param best_model_name:
    :param multi_gpu:
    :param clipping:
    :param num_thread:
301
302
    :return:
    """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
303
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
304

305
    # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
306
    # if model_name is not None:
Anthony Larcher's avatar
Anthony Larcher committed
307
308
309
310
311
    #    # Load the model
    #    logging.critical(f"*** Load model from = {model_name}")
    #    checkpoint = torch.load(model_name)
    #    model = Xtractor(speaker_number, model_yaml)
    #    model.load_state_dict(checkpoint["model_state_dict"])
Anthony Larcher's avatar
Anthony Larcher committed
312
    # else:
Anthony Larcher's avatar
Anthony Larcher committed
313
    if True:
Anthony Larcher's avatar
Anthony Larcher committed
314
315
        # Initialize a first model
        if model_yaml is None:
Anthony Larcher's avatar
Anthony Larcher committed
316
            model = Xtractor(speaker_number)
Anthony Larcher's avatar
Anthony Larcher committed
317
        else:
Anthony Larcher's avatar
Anthony Larcher committed
318
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
319

Anthony Larcher's avatar
Anthony Larcher committed
320
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
321
322
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
323
324
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
325
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
326
327

    """
Anthony Larcher's avatar
Anthony Larcher committed
328
329
330
331
    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
332
    """
Anthony Larcher's avatar
Anthony Larcher committed
333
334
335
336
    with open(dataset_yaml, "r") as fh:
        dataset_params = yaml.load(fh, Loader=yaml.FullLoader)
        df = pandas.read_csv(dataset_params["dataset_description"])
    training_df, validation_df = train_test_split(df, test_size=dataset_params["validation_ratio"])
337

Anthony Larcher's avatar
Anthony Larcher committed
338
    torch.manual_seed(dataset_params['seed'])
339
340
341
342
343
    training_set = SideSet(dataset_yaml, 
                           set_type="train", 
                           dataset_df=training_df, 
                           chunk_per_segment=dataset_params['chunk_per_segment'], 
                           overlap=dataset_params['overlap'])
Anthony Larcher's avatar
Anthony Larcher committed
344
345
346
    training_loader = DataLoader(training_set,
                                 batch_size=dataset_params["batch_size"],
                                 shuffle=True,
Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
347
                                 drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
348
                                 num_workers=num_thread)
349

Anthony Larcher's avatar
Anthony Larcher committed
350
351
352
    validation_set = SideSet(dataset_yaml, set_type="validation", dataset_df=validation_df)
    validation_loader = DataLoader(validation_set,
                                   batch_size=dataset_params["batch_size"],
Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
353
                                   drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
354
                                   num_workers=num_thread)
Anthony Larcher's avatar
Anthony Larcher committed
355
356
357
358
359

    """
    Set the training options
    """
    if type(model) is Xtractor:
Anthony Larcher's avatar
Anthony Larcher committed
360
        optimizer = torch.optim.SGD([
Anthony Larcher's avatar
Anthony Larcher committed
361
362
363
364
365
366
            {'params': model.sequence_network.parameters(),
             'weight_decay': model.sequence_network_weight_decay},
            {'params': model.before_speaker_embedding.parameters(),
             'weight_decay': model.before_speaker_embedding_weight_decay},
            {'params': model.after_speaker_embedding.parameters(),
             'weight_decay': model.after_speaker_embedding_weight_decay}],
Anthony Larcher's avatar
Anthony Larcher committed
367
            lr=lr, momentum=0.9
Anthony Larcher's avatar
Anthony Larcher committed
368
369
        )
    else:
Anthony Larcher's avatar
Anthony Larcher committed
370
        optimizer = torch.optim.SGD([
Anthony Larcher's avatar
Anthony Larcher committed
371
372
373
374
375
376
            {'params': model.module.sequence_network.parameters(),
             'weight_decay': model.module.sequence_network_weight_decay},
            {'params': model.module.before_speaker_embedding.parameters(),
             'weight_decay': model.module.before_speaker_embedding_weight_decay},
            {'params': model.module.after_speaker_embedding.parameters(),
             'weight_decay': model.module.after_speaker_embedding_weight_decay}],
Anthony Larcher's avatar
Anthony Larcher committed
377
            lr=lr, momentum=0.9
Anthony Larcher's avatar
Anthony Larcher committed
378
        )
Anthony Larcher's avatar
Anthony Larcher committed
379
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', verbose=True)
380

Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
381
    best_accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
382
    best_accuracy_epoch = 1
Anthony Larcher's avatar
Anthony Larcher committed
383
    for epoch in range(1, epochs + 1):
384
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
385
386
387
388
389
390
391
        model = train_epoch(model,
                            epoch,
                            training_loader,
                            optimizer,
                            dataset_params["log_interval"],
                            device=device,
                            clipping=clipping)
392
393

        # Add the cross validation here
Anthony Larcher's avatar
Anthony Larcher committed
394
        accuracy, val_loss = cross_validation(model, validation_loader, device=device)
395
396
397
398
        logging.critical("*** Cross validation accuracy = {} %".format(accuracy))

        # Decrease learning rate according to the scheduler policy
        scheduler.step(val_loss)
Anthony Larcher's avatar
Anthony Larcher committed
399
        print(f"Learning rate is {optimizer.param_groups[0]['lr']}")
400

Anthony Larcher's avatar
Anthony Larcher committed
401
402
403
404
        # remember best accuracy and save checkpoint
        is_best = accuracy > best_accuracy
        best_accuracy = max(accuracy, best_accuracy)

Anthony Larcher's avatar
Anthony Larcher committed
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
        if type(model) is Xtractor:
            save_checkpoint({
                'epoch': epoch,
                'model_state_dict': model.state_dict(),
                'optimizer_state_dict': optimizer.state_dict(),
                'accuracy': best_accuracy,
                'scheduler': scheduler
            }, is_best, filename=tmp_model_name+".pt", best_filename=best_model_name+'.pt')
        else:
            save_checkpoint({
                'epoch': epoch,
                'model_state_dict': model.module.state_dict(),
                'optimizer_state_dict': optimizer.state_dict(),
                'accuracy': best_accuracy,
                'scheduler': scheduler
            }, is_best, filename=tmp_model_name+".pt", best_filename=best_model_name+'.pt')
Anthony Larcher's avatar
Anthony Larcher committed
421
422
423

        if is_best:
            best_accuracy_epoch = epoch
424

Anthony Larcher's avatar
Anthony Larcher committed
425
    logging.critical(f"Best accuracy {best_accuracy * 100.} obtained at epoch {best_accuracy_epoch}")
426

Anthony Larcher's avatar
Anthony Larcher committed
427

Anthony Larcher's avatar
Anthony Larcher committed
428
def train_epoch(model, epoch, training_loader, optimizer, log_interval, device, clipping=False):
429
430
431
432
    """

    :param model:
    :param epoch:
Anthony Larcher's avatar
Anthony Larcher committed
433
    :param training_loader:
434
    :param optimizer:
Anthony Larcher's avatar
Anthony Larcher committed
435
436
437
    :param log_interval:
    :param device:
    :param clipping:
438
439
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
440
    model.train()
Anthony Larcher's avatar
Anthony Larcher committed
441
    criterion = torch.nn.CrossEntropyLoss(reduction='mean')
442
443

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
444
    for batch_idx, (data, target) in enumerate(training_loader):
445
446
447
448
449
        target = target.squeeze()
        optimizer.zero_grad()
        output = model(data.to(device))
        loss = criterion(output, target.to(device))
        loss.backward()
Anthony Larcher's avatar
Anthony Larcher committed
450
451
        if clipping:
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
452
453
454
        optimizer.step()
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()

Anthony Larcher's avatar
Anthony Larcher committed
455
        if batch_idx % log_interval == 0:
Anthony Larcher's avatar
Anthony Larcher committed
456
            batch_size = target.shape[0]
457
            logging.critical('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
Anthony Larcher's avatar
Anthony Larcher committed
458
                epoch, batch_idx + 1, training_loader.__len__(),
Anthony Larcher's avatar
Anthony Larcher committed
459
460
                100. * batch_idx / training_loader.__len__(), loss.item(),
                100.0 * accuracy.item() / ((batch_idx + 1) * batch_size)))
461
462
463
    return model


Anthony Larcher's avatar
Anthony Larcher committed
464
def cross_validation(model, validation_loader, device):
465
466
467
    """

    :param model:
Anthony Larcher's avatar
Anthony Larcher committed
468
469
    :param validation_loader:
    :param device:
470
471
472
473
474
    :return:
    """
    model.eval()

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
475
    loss = 0.0
476
    criterion = torch.nn.CrossEntropyLoss()
Anthony Larcher's avatar
Anthony Larcher committed
477
478
479
480
481
482
    with torch.no_grad():
        for batch_idx, (data, target) in enumerate(validation_loader):
            batch_size = target.shape[0]
            target = target.squeeze()
            output = model(data.to(device))
            accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()
483

Anthony Larcher's avatar
Anthony Larcher committed
484
485
            loss += criterion(output, target.to(device))
    
Anthony Larcher's avatar
Anthony Larcher committed
486
487
488
489
    return 100. * accuracy.cpu().numpy() / ((batch_idx + 1) * batch_size), \
           loss.cpu().numpy() / ((batch_idx + 1) * batch_size)


490
def extract_embeddings(idmap, speaker_number, model_filename, model_yaml, data_root_name , device):
Anthony Larcher's avatar
Anthony Larcher committed
491
492
493
494
495
496
497
498
499
500

    # Create dataset to load the data
    dataset = IdMapSet(data_root_name, idmap_name)

    # Load the model
    checkpoint = torch.load(model_filename)
    model = Xtractor(speaker_number, model_archi=model_yaml)
    model.load_state_dict(checkpoint["model_state_dict"])
    model.eval()
    model.to(device)
501

Anthony Larcher's avatar
Anthony Larcher committed
502
503
504
    # Get the size of embeddings to extract
    name = list(model.before_speaker_embedding.state_dict().keys())[-1].split('.')[0] + '.weight'
    emb_size = model.before_speaker_embedding.state_dict()[name].shape[0]
505

Anthony Larcher's avatar
Anthony Larcher committed
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
    # Create the StatServer
    embeddings = sidekit.StatServer()
    embeddings.modelset = idmap.leftids
    embeddings.segset = idmap.rightids
    embeddings.start = idmap.start
    embeddings.stop = idmap.stop
    embeddings.stat0 = numpy.ones((embeddings.modelset.shape[0], 1))
    embeddings.stat1 = numpy.ones((embeddings.modelset.shape[0], emb_size))

    # Process the data
    with torch.no_grad():
        for idx, (data, mod, seg) in tqdm(enumerate(dataset)):
            vec = model(data.to(device), is_eval=True)
            current_idx = numpy.argwhere(numpy.logical_and(im.leftids == mod, im.rightids == seg))[0][0]
            embeddings.stat1[current_idx, :] = vec.detach().cpu()

    return embeddings


def extract_idmap(args, segment_indices, fs_params, idmap_name, output_queue):
Anthony Larcher's avatar
Anthony Larcher committed
526
    """
Anthony Larcher's avatar
Anthony Larcher committed
527
528
    Function that takes a model and an idmap and extract all x-vectors based on this model
    and return a StatServer containing the x-vectors
Anthony Larcher's avatar
Anthony Larcher committed
529
530
531
532
533
534
535

    :param args:
    :param segment_indices:
    :param fs_params:
    :param idmap_name:
    :param output_queue:
    :return:
Anthony Larcher's avatar
Anthony Larcher committed
536
    """
537
    # device = torch.device("cuda:{}".format(device_ID))
Anthony Larcher's avatar
Anthony Larcher committed
538
    device = torch.device('cpu')
Anthony Larcher's avatar
Anthony Larcher committed
539
540
541
542
543
544
545
546
547
548
549
550
551

    # Create the dataset
    tmp_idmap = IdMap(idmap_name)
    idmap = IdMap()
    idmap.leftids = tmp_idmap.leftids[segment_indices]
    idmap.rightids = tmp_idmap.rightids[segment_indices]
    idmap.start = tmp_idmap.start[segment_indices]
    idmap.stop = tmp_idmap.stop[segment_indices]

    segment_loader = StatDataset(idmap, fs_params)

    # Load the model
    model_file_name = '/'.join([args.model_path, args.model_name])
Anthony Larcher's avatar
Anthony Larcher committed
552
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
553
554
555
556
557
558
559
560
    model.load_state_dict(torch.load(model_file_name))
    model.eval()

    # Get the size of embeddings
    emb_a_size = model.seg_lin0.weight.data.shape[0]
    emb_b_size = model.seg_lin1.weight.data.shape[0]

    # Create a Tensor to store all x-vectors on the GPU
Anthony Larcher's avatar
Anthony Larcher committed
561
562
563
564
565
566
    emb_1 = numpy.zeros((idmap.leftids.shape[0], emb_a_size)).astype(numpy.float32)
    emb_2 = numpy.zeros((idmap.leftids.shape[0], emb_b_size)).astype(numpy.float32)
    emb_3 = numpy.zeros((idmap.leftids.shape[0], emb_b_size)).astype(numpy.float32)
    emb_4 = numpy.zeros((idmap.leftids.shape[0], emb_b_size)).astype(numpy.float32)
    emb_5 = numpy.zeros((idmap.leftids.shape[0], emb_b_size)).astype(numpy.float32)
    emb_6 = numpy.zeros((idmap.leftids.shape[0], emb_b_size)).astype(numpy.float32)
Anthony Larcher's avatar
Anthony Larcher committed
567
568
569
570
571
572

    # Send on selected device
    model.to(device)

    # Loop to extract all x-vectors
    for idx, (model_id, segment_id, data) in enumerate(segment_loader):
Anthony Larcher's avatar
Anthony Larcher committed
573
        logging.critical('Process file {}, [{} / {}]'.format(segment_id, idx, segment_loader.__len__()))
Anthony Larcher's avatar
Anthony Larcher committed
574

Anthony Larcher's avatar
Anthony Larcher committed
575
576
577
        if list(data.shape)[2] < 20:
            pass
        else:
Anthony Larcher's avatar
Anthony Larcher committed
578
579
580
581
582
583
584
            seg_1, seg_2, seg_3, seg_4, seg_5, seg_6 = model.extract(data.to(device))
            emb_1[idx, :] = seg_1.detach().cpu()
            emb_2[idx, :] = seg_2.detach().cpu()
            emb_3[idx, :] = seg_3.detach().cpu()
            emb_4[idx, :] = seg_4.detach().cpu()
            emb_5[idx, :] = seg_5.detach().cpu()
            emb_6[idx, :] = seg_6.detach().cpu()
Anthony Larcher's avatar
Anthony Larcher committed
585

Anthony Larcher's avatar
Anthony Larcher committed
586
    output_queue.put((segment_indices, emb_1, emb_2, emb_3, emb_4, emb_5, emb_6))
Anthony Larcher's avatar
Anthony Larcher committed
587
588


Anthony Larcher's avatar
Anthony Larcher committed
589
def extract_parallel(args, fs_params):
590
591
592
593
594
595
    """

    :param args:
    :param fs_params:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
596
597
598
    emb_a_size = 512
    emb_b_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
599
    idmap = IdMap(args.idmap)
Anthony Larcher's avatar
Anthony Larcher committed
600

Anthony Larcher's avatar
Anthony Larcher committed
601
602
603
604
605
606
607
608
609
610
611
612
613
    x_server_1 = StatServer(idmap, 1, emb_a_size)
    x_server_2 = StatServer(idmap, 1, emb_b_size)
    x_server_3 = StatServer(idmap, 1, emb_b_size)
    x_server_4 = StatServer(idmap, 1, emb_b_size)
    x_server_5 = StatServer(idmap, 1, emb_b_size)
    x_server_6 = StatServer(idmap, 1, emb_b_size)

    x_server_1.stat0 = numpy.ones(x_server_1.stat0.shape)
    x_server_2.stat0 = numpy.ones(x_server_2.stat0.shape)
    x_server_3.stat0 = numpy.ones(x_server_3.stat0.shape)
    x_server_4.stat0 = numpy.ones(x_server_4.stat0.shape)
    x_server_5.stat0 = numpy.ones(x_server_5.stat0.shape)
    x_server_6.stat0 = numpy.ones(x_server_6.stat0.shape)
Anthony Larcher's avatar
Anthony Larcher committed
614
615
616

    # Split the indices
    mega_batch_size = idmap.leftids.shape[0] // args.num_processes
Anthony Larcher's avatar
Anthony Larcher committed
617
618
619

    logging.critical("Number of sessions to process: {}".format(idmap.leftids.shape[0]))

Anthony Larcher's avatar
Anthony Larcher committed
620
621
622
    segment_idx = []
    for ii in range(args.num_processes):
        segment_idx.append(
Anthony Larcher's avatar
Anthony Larcher committed
623
624
625
626
            numpy.arange(ii * mega_batch_size, numpy.min([(ii + 1) * mega_batch_size, idmap.leftids.shape[0]])))

    for idx, si in enumerate(segment_idx):
        logging.critical("Number of session on process {}: {}".format(idx, len(si)))
Anthony Larcher's avatar
Anthony Larcher committed
627
628
629
630
631
632
633

    # Extract x-vectors in parallel
    output_queue = mp.Queue()

    processes = []
    for rank in range(args.num_processes):
        p = mp.Process(target=extract_idmap,
Anthony Larcher's avatar
Anthony Larcher committed
634
                       args=(args, rank, segment_idx[rank], fs_params, args.idmap, output_queue)
Anthony Larcher's avatar
Anthony Larcher committed
635
636
637
638
639
640
641
                       )
        # We first train the model across `num_processes` processes
        p.start()
        processes.append(p)

    # Get the x-vectors and fill the StatServer
    for ii in range(args.num_processes):
Anthony Larcher's avatar
Anthony Larcher committed
642
643
644
645
646
647
648
        indices, seg_1, seg_2, seg_3, seg_4, seg_5, seg_6 = output_queue.get()
        x_server_1.stat1[indices, :] = seg_1
        x_server_2.stat1[indices, :] = seg_2
        x_server_3.stat1[indices, :] = seg_3
        x_server_4.stat1[indices, :] = seg_4
        x_server_5.stat1[indices, :] = seg_5
        x_server_6.stat1[indices, :] = seg_6
Anthony Larcher's avatar
Anthony Larcher committed
649
650
651
652

    for p in processes:
        p.join()

Anthony Larcher's avatar
Anthony Larcher committed
653
    return x_server_1, x_server_2, x_server_3, x_server_4, x_server_5, x_server_6
Anthony Larcher's avatar
Anthony Larcher committed
654