xvector.py 22.9 KB
Newer Older
Anthony Larcher's avatar
Anthony Larcher committed
1
# coding: utf-8 -*-
Anthony Larcher's avatar
Anthony Larcher committed
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
#
# This file is part of SIDEKIT.
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is free software: you can redistribute it and/or modify
# it under the terms of the GNU LLesser General Public License as
# published by the Free Software Foundation, either version 3 of the License,
# or (at your option) any later version.
#
# SIDEKIT is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU Lesser General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with SIDEKIT.  If not, see <http://www.gnu.org/licenses/>.

"""
Anthony Larcher's avatar
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
from tqdm import tqdm
Anthony Larcher's avatar
Anthony Larcher committed
49

Anthony Larcher's avatar
Anthony Larcher committed
50
51
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
52
__copyright__ = "Copyright 2015-2020 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
53
54
55
56
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
57
58


59
def get_lr(optimizer):
Anthony Larcher's avatar
Anthony Larcher committed
60
61
62
63
64
    """

    :param optimizer:
    :return:
    """
65
66
67
68
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
69
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
70
71
72
73
74
75
76
77
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
78
79
80
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
81
82
83


class Xtractor(torch.nn.Module):
84
85
86
    """
    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
87

Anthony Larcher's avatar
Anthony Larcher committed
88
    def __init__(self, speaker_number, model_archi=None):
Anthony Larcher's avatar
Anthony Larcher committed
89
90
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
91
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
92
        """
Anthony Larcher's avatar
Anthony Larcher committed
93
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
94
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
95
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
96

Anthony Larcher's avatar
Anthony Larcher committed
97
        if model_archi is None:
Anthony Larcher's avatar
Anthony Larcher committed
98
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
99
100
101
            self.activation = torch.nn.LeakyReLU(0.2)

            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
102

Anthony Larcher's avatar
xv    
Anthony Larcher committed
103
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
104
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
105
106
107
108
109
110
111
112
                ("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
113
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
114
115
                ("activation4", torch.nn.LeakyReLU(0.2)),
                ("norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
116
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
117
118
119
120
                ("activation5", torch.nn.LeakyReLU(0.2)),
                ("norm5", torch.nn.BatchNorm1d(1536))
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
121
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
122
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
123
124
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
125
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
126
127
                ("activation6", torch.nn.LeakyReLU(0.2)),
                ("norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
128
                ("dropout6", torch.nn.Dropout(p=0.05)),
Anthony Larcher's avatar
Anthony Larcher committed
129
                ("linear7", torch.nn.Linear(512, 512)),
Anthony Larcher's avatar
Anthony Larcher committed
130
131
                ("activation7", torch.nn.LeakyReLU(0.2)),
                ("norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
132
                ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
Anthony Larcher's avatar
Anthony Larcher committed
133
134
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
135
136
137
138
            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
139
140
        else:
            # Load Yaml configuration
Anthony Larcher's avatar
Anthony Larcher committed
141
            with open(model_archi, 'r') as fh:
Anthony Larcher's avatar
Anthony Larcher committed
142
143
                cfg = yaml.load(fh, Loader=yaml.FullLoader)

Anthony Larcher's avatar
Anthony Larcher committed
144
145
146
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
147
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
148
149
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
150
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
151
152
153
154
155
156
157
158
159
160
161
162
                        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
163
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
164
165

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
166
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
167
            """
Anthony Larcher's avatar
Anthony Larcher committed
168
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
169
170
171
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
            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
188
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
189
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
190
191
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
192
193
194
195
196
197
198
199
                    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
200
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
201
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
202

Anthony Larcher's avatar
Anthony Larcher committed
203
204
205
            """
            Prepapre last part of the network (after pooling)
            """
Anthony Larcher's avatar
Anthony Larcher committed
206
207
            # Create sequential object for the second part of the network
            input_size = input_size * 2
Anthony Larcher's avatar
xv    
Anthony Larcher committed
208
209
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
210
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
211
212
                    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
213
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
214
215
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
216
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
217
218

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
219
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
220
221

                elif k.startswith('norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
222
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
223
224

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
225
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
226

Anthony Larcher's avatar
Anthony Larcher committed
227
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
228
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
229
230
231
232
233

            # 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
234
235
                    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
236
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
237
238
                        after_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
239
                        input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
240
241
242
243
244
245
246
247

                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
248
                    after_embedding_layers.append((k, torch.nn.Dropout(p=cfg["after_embedding"][k])))
Anthony Larcher's avatar
Anthony Larcher committed
249

Anthony Larcher's avatar
Anthony Larcher committed
250
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict(after_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
251
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
252

Anthony Larcher's avatar
Anthony Larcher committed
253
    def forward(self, x, is_eval=False):
254
255
256
        """

        :param x:
Anthony Larcher's avatar
Anthony Larcher committed
257
        :param is_eval:
258
259
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
260
261
262
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
263
        x = self.sequence_network(x)
264

Anthony Larcher's avatar
Anthony Larcher committed
265
266
267
268
269
270
271
272
        # 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
273

Anthony Larcher's avatar
Anthony Larcher committed
274
275
        x = self.after_speaker_embedding(x)
        return x
Anthony Larcher's avatar
Anthony Larcher committed
276

Anthony Larcher's avatar
Anthony Larcher committed
277

Anthony Larcher's avatar
Anthony Larcher committed
278
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
279
           dataset_yaml,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
280
           epochs=100,
Anthony Larcher's avatar
Anthony Larcher committed
281
           lr=0.01,
Anthony Larcher's avatar
Anthony Larcher committed
282
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
283
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
284
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
285
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
286
           multi_gpu=True,
287
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
288
           opt='sgd',
Anthony Larcher's avatar
Anthony Larcher committed
289
           num_thread=1):
290
291
    """

Anthony Larcher's avatar
Anthony Larcher committed
292
293
294
295
296
297
298
299
300
301
302
    :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:
303
304
    :return:
    """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
305
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
306

307
    # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
308
309
310
311
312
313
314
    if model_name is not None:
        # 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"])
    else:
Anthony Larcher's avatar
Anthony Larcher committed
315
316
        # Initialize a first model
        if model_yaml is None:
Anthony Larcher's avatar
Anthony Larcher committed
317
            model = Xtractor(speaker_number)
Anthony Larcher's avatar
Anthony Larcher committed
318
        else:
Anthony Larcher's avatar
Anthony Larcher committed
319
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
320

Anthony Larcher's avatar
Anthony Larcher committed
321
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
322
323
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
324
325
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
326
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
327
328

    """
Anthony Larcher's avatar
Anthony Larcher committed
329
330
331
332
    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
333
    """
Anthony Larcher's avatar
Anthony Larcher committed
334
335
336
337
    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"])
338

Anthony Larcher's avatar
Anthony Larcher committed
339
    torch.manual_seed(dataset_params['seed'])
340
341
342
    training_set = SideSet(dataset_yaml, 
                           set_type="train", 
                           dataset_df=training_df, 
Anthony Larcher's avatar
Anthony Larcher committed
343
344
                           chunk_per_segment=dataset_params['train']['chunk_per_segment'], 
                           overlap=dataset_params['train']['overlap'])
Anthony Larcher's avatar
Anthony Larcher committed
345
346
347
    training_loader = DataLoader(training_set,
                                 batch_size=dataset_params["batch_size"],
                                 shuffle=True,
Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
348
                                 drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
349
                                 num_workers=num_thread)
350

Anthony Larcher's avatar
Anthony Larcher committed
351
352
353
    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
354
                                   drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
355
                                   num_workers=num_thread)
Anthony Larcher's avatar
Anthony Larcher committed
356
357
358
359

    """
    Set the training options
    """
Anthony Larcher's avatar
Anthony Larcher committed
360
361
    if opt == 'sgd':
        _optimizer = torch.optim.SGD
Anthony Larcher's avatar
Anthony Larcher committed
362
        _options = {'lr': lr, 'momentum': 0.9}
Anthony Larcher's avatar
Anthony Larcher committed
363
364
    elif opt == 'adam':
        _optimizer = torch.optim.Adam
Anthony Larcher's avatar
Anthony Larcher committed
365
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
366
367
    elif opt == 'rmsprop':
        _optimizer = torch.optim.RMSprop
Anthony Larcher's avatar
Anthony Larcher committed
368
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
369

Anthony Larcher's avatar
Anthony Larcher committed
370
    if type(model) is Xtractor:
Anthony Larcher's avatar
Anthony Larcher committed
371
        optimizer = _optimizer([
Anthony Larcher's avatar
Anthony Larcher committed
372
373
374
375
376
377
            {'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
378
            **_options
Anthony Larcher's avatar
Anthony Larcher committed
379
380
        )
    else:
Anthony Larcher's avatar
Anthony Larcher committed
381
        optimizer = _optimizer([
Anthony Larcher's avatar
Anthony Larcher committed
382
383
384
385
386
387
            {'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
388
            **_options
Anthony Larcher's avatar
Anthony Larcher committed
389
        )
Anthony Larcher's avatar
Anthony Larcher committed
390
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', verbose=True)
391

Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
392
    best_accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
393
    best_accuracy_epoch = 1
Anthony Larcher's avatar
Anthony Larcher committed
394
    for epoch in range(1, epochs + 1):
395
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
396
397
398
399
400
401
402
        model = train_epoch(model,
                            epoch,
                            training_loader,
                            optimizer,
                            dataset_params["log_interval"],
                            device=device,
                            clipping=clipping)
403
404

        # Add the cross validation here
Anthony Larcher's avatar
Anthony Larcher committed
405
        accuracy, val_loss = cross_validation(model, validation_loader, device=device)
406
407
408
409
        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
410
        print(f"Learning rate is {optimizer.param_groups[0]['lr']}")
411

Anthony Larcher's avatar
Anthony Larcher committed
412
413
414
415
        # remember best accuracy and save checkpoint
        is_best = accuracy > best_accuracy
        best_accuracy = max(accuracy, best_accuracy)

Anthony Larcher's avatar
Anthony Larcher committed
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
        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
432
433
434

        if is_best:
            best_accuracy_epoch = epoch
435

Anthony Larcher's avatar
Anthony Larcher committed
436
    logging.critical(f"Best accuracy {best_accuracy * 100.} obtained at epoch {best_accuracy_epoch}")
437

Anthony Larcher's avatar
Anthony Larcher committed
438

Anthony Larcher's avatar
Anthony Larcher committed
439
def train_epoch(model, epoch, training_loader, optimizer, log_interval, device, clipping=False):
440
441
442
443
    """

    :param model:
    :param epoch:
Anthony Larcher's avatar
Anthony Larcher committed
444
    :param training_loader:
445
    :param optimizer:
Anthony Larcher's avatar
Anthony Larcher committed
446
447
448
    :param log_interval:
    :param device:
    :param clipping:
449
450
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
451
    model.train()
Anthony Larcher's avatar
Anthony Larcher committed
452
    criterion = torch.nn.CrossEntropyLoss(reduction='mean')
453
454

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
455
    for batch_idx, (data, target) in enumerate(training_loader):
456
457
458
459
460
        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
461
462
        if clipping:
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
463
464
465
        optimizer.step()
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()

Anthony Larcher's avatar
Anthony Larcher committed
466
        if batch_idx % log_interval == 0:
Anthony Larcher's avatar
Anthony Larcher committed
467
            batch_size = target.shape[0]
468
            logging.critical('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
Anthony Larcher's avatar
Anthony Larcher committed
469
                epoch, batch_idx + 1, training_loader.__len__(),
Anthony Larcher's avatar
Anthony Larcher committed
470
471
                100. * batch_idx / training_loader.__len__(), loss.item(),
                100.0 * accuracy.item() / ((batch_idx + 1) * batch_size)))
472
473
474
    return model


Anthony Larcher's avatar
Anthony Larcher committed
475
def cross_validation(model, validation_loader, device):
476
477
478
    """

    :param model:
Anthony Larcher's avatar
Anthony Larcher committed
479
480
    :param validation_loader:
    :param device:
481
482
483
484
485
    :return:
    """
    model.eval()

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
486
    loss = 0.0
487
    criterion = torch.nn.CrossEntropyLoss()
Anthony Larcher's avatar
Anthony Larcher committed
488
489
490
491
492
493
    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()
494

Anthony Larcher's avatar
Anthony Larcher committed
495
496
            loss += criterion(output, target.to(device))
    
Anthony Larcher's avatar
Anthony Larcher committed
497
498
499
500
    return 100. * accuracy.cpu().numpy() / ((batch_idx + 1) * batch_size), \
           loss.cpu().numpy() / ((batch_idx + 1) * batch_size)


Anthony Larcher's avatar
Anthony Larcher committed
501
502
503
504
505
506
507
508
def extract_embeddings(idmap_name,
                       speaker_number,
                       model_filename,
                       model_yaml,
                       data_root_name ,
                       device,
                       file_extension="wav",
                       transform_pipeline=None):
Anthony Larcher's avatar
Anthony Larcher committed
509

510
511
512
513
514
    if isinstance(idmap_name, sidekit.bosaris.idmap.IdMap):
        idmap = idmap_name
    else:
        idmap = IdMap(idmap_name)

Anthony Larcher's avatar
Anthony Larcher committed
515
    # Create dataset to load the data
Anthony Larcher's avatar
Anthony Larcher committed
516
517
518
519
    dataset = IdMapSet(idmap_name=idmap_name,
                       data_root_path=data_root_name,
                       file_extension=file_extension,
                       transform_pipeline=transform_pipeline)
Anthony Larcher's avatar
Anthony Larcher committed
520
521

    # Load the model
522
523
524
525
526
527
528
    if isinstance(model_filename, str):
        checkpoint = torch.load(model_filename)
        model = Xtractor(speaker_number, model_archi=model_yaml)
        model.load_state_dict(checkpoint["model_state_dict"])
    else:
        model = model_filename

Anthony Larcher's avatar
Anthony Larcher committed
529
530
    model.eval()
    model.to(device)
531

Anthony Larcher's avatar
Anthony Larcher committed
532
533
534
    # 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]
Anthony Larcher's avatar
Anthony Larcher committed
535
    
Anthony Larcher's avatar
Anthony Larcher committed
536
    # Create the StatServer
Anthony Larcher's avatar
Anthony Larcher committed
537
    embeddings = StatServer()
Anthony Larcher's avatar
Anthony Larcher committed
538
539
540
541
542
543
    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))
Anthony Larcher's avatar
Anthony Larcher committed
544
    
Anthony Larcher's avatar
Anthony Larcher committed
545
546
    # Process the data
    with torch.no_grad():
Anthony Larcher's avatar
Anthony Larcher committed
547
        for idx in tqdm(range(len(dataset))):
Anthony Larcher's avatar
Anthony Larcher committed
548
            data, mod, seg, _, __ = dataset[idx]
Anthony Larcher's avatar
Anthony Larcher committed
549
550
            vec = model(data[None, :, :].to(device), is_eval=True)
            current_idx = numpy.argwhere(numpy.logical_and(idmap.leftids == mod, idmap.rightids == seg))[0][0]
Anthony Larcher's avatar
Anthony Larcher committed
551
552
553
554
555
            embeddings.stat1[current_idx, :] = vec.detach().cpu()

    return embeddings


Anthony Larcher's avatar
Anthony Larcher committed
556
557
558
559
560
561
562
563
564
565
566
567
def extract_sliding_embedding(idmap_name,
                              window_length,
                              sample_rate,
                              overlap,
                              speaker_number,
                              model_filename,
                              model_yaml,
                              data_root_name ,
                              device,
                              file_extension="wav",
                              transform_pipeline=None):

568
569
570
571
572
573
574
575
576
577
578
579
580
581

    # From the original IdMap, create the new one to extract x-vectors
    input_idmap = IdMap(idmap_name)

    # Create temporary lists
    nb_chunks = 0
    model_names = []
    segment_names = []
    starts = []
    stops = []
    for mod, seg, start, stop in zip(input_idmap.leftids, input_idmap.rightids, input_idmap.start, input_idmap.stop):
        # Compute the number of chunks to process
        chunk_starts = numpy.arange(start,
                                    stop - int(sample_rate * window_length),
Anthony Larcher's avatar
Anthony Larcher committed
582
                                    int(sample_rate * (window_length - overlap)))
583
584
585
586
587
588
589
590
591
592

        # Create a numpy array to store the current x-vectors
        model_names.append(numpy.array([mod, ] * chunk_starts.shape[0]))
        segment_names.append(numpy.array([seg, ] * chunk_starts.shape[0]))
        starts.append(chunk_starts)
        stops.append(chunk_starts + sample_rate * window_length)

        nb_chunks += len(chunk_starts)

    sliding_idmap = IdMap()
Anthony Larcher's avatar
Anthony Larcher committed
593
594
595
596
    sliding_idmap.leftids = numpy.hstack(model_names)
    sliding_idmap.rightids = numpy.hstack(segment_names)
    sliding_idmap.start = numpy.hstack(starts)
    sliding_idmap.stop = numpy.hstack(stops)
597
    assert sliding_idmap.validate()
Anthony Larcher's avatar
Anthony Larcher committed
598

Anthony Larcher's avatar
Anthony Larcher committed
599
600
601
602
603
604
    embeddings = extract_embeddings(sliding_idmap,
                                 speaker_number,
                                 model_filename,
                                 model_yaml,
                                 data_root_name,
                                 device,
Anthony Larcher's avatar
Anthony Larcher committed
605
606
                                 file_extension=file_extension,
                                 transform_pipeline=transform_pipeline)
Anthony Larcher's avatar
Anthony Larcher committed
607
608

    return embeddings