xvector.py 30.9 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
41
from .xsets import XvectorMultiDataset, StatDataset, VoxDataset, SideSet
from .xsets import FrequencyMask, CMVN, TemporalMask, MFCC
Anthony Larcher's avatar
Anthony Larcher committed
42
43
from ..bosaris import IdMap
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
44
from torch.utils.data import DataLoader
Anthony Larcher's avatar
Anthony Larcher committed
45
from sklearn.model_selection import train_test_split
Anthony Larcher's avatar
Anthony Larcher committed
46

Anthony Larcher's avatar
Anthony Larcher committed
47
48
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
49
__copyright__ = "Copyright 2015-2020 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
50
51
52
53
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
54
55


Anthony Larcher's avatar
Anthony Larcher committed
56
# logging.basicConfig(stream=sys.stdout, level=logging.INFO)
Anthony Larcher's avatar
Anthony Larcher committed
57
58


59
60
61
62
63
def get_lr(optimizer):
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
64
65
66
67
68
69
70
def split_file_list(batch_files, num_processes):
    # Cut the list of files into args.num_processes lists of files
    batch_sub_lists = [[]] * num_processes
    x = [ii for ii in range(len(batch_files))]
    for ii in range(num_processes):
        batch_sub_lists[ii - 1] = [batch_files[z + ii] for z in x[::num_processes] if (z + ii) < len(batch_files)]
    return batch_sub_lists
Anthony Larcher's avatar
Anthony Larcher committed
71
72
73


class Xtractor(torch.nn.Module):
74
75
76
    """
    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
77

Anthony Larcher's avatar
Anthony Larcher committed
78
    def __init__(self, speaker_number, model_archi=None):
Anthony Larcher's avatar
Anthony Larcher committed
79
80
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
81
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
82
        """
Anthony Larcher's avatar
Anthony Larcher committed
83
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
84
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
85
       self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
86

Anthony Larcher's avatar
Anthony Larcher committed
87
        if model_archi is None:
Anthony Larcher's avatar
Anthony Larcher committed
88
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
89
90
            self.activation = torch.nn.ReLU()

Anthony Larcher's avatar
xv    
Anthony Larcher committed
91
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
92
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
93
94
95
96
97
98
99
100
                ("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
101
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
102
103
                ("activation4", torch.nn.LeakyReLU(0.2)),
                ("norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
104
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
105
106
107
108
                ("activation5", torch.nn.LeakyReLU(0.2)),
                ("norm5", torch.nn.BatchNorm1d(1536))
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
109
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
110
                ("linear6", torch.nn.Linear(1536, 512))
Anthony Larcher's avatar
Anthony Larcher committed
111
112
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
113
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
114
115
                ("activation6", torch.nn.LeakyReLU(0.2)),
                ("norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
116
                ("linear7", torch.nn.Linear(512, 512)),
Anthony Larcher's avatar
Anthony Larcher committed
117
118
                ("activation7", torch.nn.LeakyReLU(0.2)),
                ("norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
119
                ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
Anthony Larcher's avatar
Anthony Larcher committed
120
121
122
123
            ]))

        else:
            # Load Yaml configuration
Anthony Larcher's avatar
Anthony Larcher committed
124
            with open(model_archi, 'r') as fh:
Anthony Larcher's avatar
Anthony Larcher committed
125
126
                cfg = yaml.load(fh, Loader=yaml.FullLoader)

Anthony Larcher's avatar
Anthony Larcher committed
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
            """
            Prepare Preprocessor
            """
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
                    self.sincnet = SincNet(
                        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"]
                    )
                    self.feature_size = self.sincnet_.dimension

            """
            Prepapre sequence network
            """
Anthony Larcher's avatar
Anthony Larcher committed
150
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
151
152
153
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
            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
170
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
171
172
173
174
175
176
177
178
179
180
181
                                                                cfg["segmental"][k]["output_channels"],
                                                                cfg["segmental"][k]["kernel_size"],
                                                                cfg["segmental"][k]["dilation"])))
                    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
182
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
183
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
184

Anthony Larcher's avatar
Anthony Larcher committed
185
186
187
            """
            Prepapre last part of the network (after pooling)
            """
Anthony Larcher's avatar
Anthony Larcher committed
188
189
            # Create sequential object for the second part of the network
            input_size = input_size * 2
Anthony Larcher's avatar
xv    
Anthony Larcher committed
190
191
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
192
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
193
194
                    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
195
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
196
197
                        before_embedding_layers.append((k, torch.nn.Linear(input_size, cfg["before_embedding"][k]["output"])))
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
198
199

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
200
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
201
202

                elif k.startswith('norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
203
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
204
205

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
206
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
207

Anthony Larcher's avatar
Anthony Larcher committed
208
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
209
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
210
211
212
213
214

            # 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
215
216
                    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
217
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
218
219
                        after_embedding_layers.append((k, torch.nn.Linear(input_size, cfg["after_embedding"][k]["output"])))
                        input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
220
221
222
223
224
225
226
227

                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
228
                    after_embedding_layers.append((k, torch.nn.Dropout(p=cfg["after_embedding"][k])))
Anthony Larcher's avatar
Anthony Larcher committed
229

Anthony Larcher's avatar
Anthony Larcher committed
230
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict(after_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
231
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
232

Anthony Larcher's avatar
Anthony Larcher committed
233
    def forward(self, x, is_eval=False):
234
235
236
237
238
        """

        :param x:
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
239
        x = self.sequence_network(x)
240

Anthony Larcher's avatar
Anthony Larcher committed
241
242
243
244
245
246
247
248
        # 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
249

Anthony Larcher's avatar
Anthony Larcher committed
250
251
        x = self.after_speaker_embedding(x)
        return x
Anthony Larcher's avatar
Anthony Larcher committed
252

Anthony Larcher's avatar
Anthony Larcher committed
253

Anthony Larcher's avatar
Anthony Larcher committed
254
255
256
257
258
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)

Anthony Larcher's avatar
minor    
Anthony Larcher committed
259

Anthony Larcher's avatar
Anthony Larcher committed
260
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
261
           dataset_yaml,
Anthony Larcher's avatar
Anthony Larcher committed
262
263
           epochs=10,
           lr=0.01,
Anthony Larcher's avatar
Anthony Larcher committed
264
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
265
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
266
           num_thread=1):
267
268
269
270
271
272
    """
    Initialize and train an x-vector on a single GPU

    :param args:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
273
274
    device = torch.device("cuda:0")

275
    # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
276
    if model_name is not None:
277
        # Load the model
Anthony Larcher's avatar
Anthony Larcher committed
278
279
        logging.critical(f"*** Load model from = {model_name}")
        checkpoint = torch.load(model_name)
Anthony Larcher's avatar
Anthony Larcher committed
280
        model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
281
        model.load_state_dict(checkpoint["model_state_dict"])
282
    else:
Anthony Larcher's avatar
Anthony Larcher committed
283
284
        # Initialize a first model
        if model_yaml is None:
Anthony Larcher's avatar
Anthony Larcher committed
285
            model = Xtractor(speaker_number)
Anthony Larcher's avatar
Anthony Larcher committed
286
        else:
Anthony Larcher's avatar
Anthony Larcher committed
287
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
288
289

    model.train()
Anthony Larcher's avatar
Anthony Larcher committed
290
291
292
293
294

    if torch.cuda.device_count() > 1:
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)

Anthony Larcher's avatar
Anthony Larcher committed
295
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
296
297

    """
Anthony Larcher's avatar
Anthony Larcher committed
298
299
300
301
    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
302
    """
Anthony Larcher's avatar
Anthony Larcher committed
303
304
305
306
    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"])
307

Anthony Larcher's avatar
Anthony Larcher committed
308
    torch.manual_seed(dataset_params['seed'])
Anthony Larcher's avatar
Anthony Larcher committed
309
310
311
312
313
    training_set = SideSet(dataset_yaml, set_type="train", dataset_df=training_df)
    training_loader = DataLoader(training_set,
                                 batch_size=dataset_params["batch_size"],
                                 shuffle=True,
                                 num_workers=num_thread)
314

Anthony Larcher's avatar
Anthony Larcher committed
315
316
317
318
    validation_set = SideSet(dataset_yaml, set_type="validation", dataset_df=validation_df)
    validation_loader = DataLoader(validation_set,
                                   batch_size=dataset_params["batch_size"],
                                   num_workers=num_thread)
Anthony Larcher's avatar
Anthony Larcher committed
319

Anthony Larcher's avatar
Anthony Larcher committed
320
321
    print(f"Length of training: {len(training_set)}, length of validation: {len(validation_set)}")

Anthony Larcher's avatar
Anthony Larcher committed
322
323
324
325
326
    """
    Set the training options
    """
    if type(model) is Xtractor:
        optimizer = torch.optim.SGD([
Anthony Larcher's avatar
Anthony Larcher committed
327
328
329
330
331
332
            {'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
333
334
335
336
            lr=lr, momentum=0.9
        )
    else:
        optimizer = torch.optim.SGD([
Anthony Larcher's avatar
Anthony Larcher committed
337
338
339
340
341
342
            {'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
343
344
            lr=lr, momentum=0.9
        )
345
346
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min')

Anthony Larcher's avatar
Anthony Larcher committed
347
    for epoch in range(1, epochs + 1):
348
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
349
        model = train_epoch(model, epoch, training_loader, optimizer, dataset_params["log_interval"], device=device)
350
351

        # Add the cross validation here
Anthony Larcher's avatar
Anthony Larcher committed
352
        accuracy, val_loss = cross_validation(model, validation_loader, device=device)
353
354
355
356
357
        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
358
359
360
361
362
363
364
365
366
367
368
369
370
371
        # remember best accuracy and save checkpoint
        is_best = accuracy > best_accuracy
        best_accuracy = max(accuracy, best_accuracy)

        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=output_model_name+'.pt')

        if is_best:
            best_accuracy_epoch = epoch
372

Anthony Larcher's avatar
Anthony Larcher committed
373
    logging.critical(f"Best accuracy {best_accuracy * 100.} obtained at epoch {best_accuracy_epoch}")
374

Anthony Larcher's avatar
Anthony Larcher committed
375
def train_epoch(model, epoch, training_loader, optimizer, log_interval, device):
376
377
378
379
380
381
382
383
384
385
386
387
388
    """

    :param model:
    :param epoch:
    :param train_seg_df:
    :param speaker_dict:
    :param optimizer:
    :param args:
    :return:
    """
    criterion = torch.nn.CrossEntropyLoss()

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
389
    for batch_idx, (data, target) in enumerate(training_loader):
390
391
392
393
394
395
396
397
        target = target.squeeze()
        optimizer.zero_grad()
        output = model(data.to(device))
        loss = criterion(output, target.to(device))
        loss.backward()
        optimizer.step()
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()

Anthony Larcher's avatar
Anthony Larcher committed
398
        if batch_idx % log_interval == 0:
Anthony Larcher's avatar
Anthony Larcher committed
399
            batch_size = target.shape[0]
400
            logging.critical('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
Anthony Larcher's avatar
Anthony Larcher committed
401
402
                epoch, batch_idx + 1, training_loader.__len__(),
                       100. * batch_idx / training_loader.__len__(), loss.item(),
Anthony Larcher's avatar
Anthony Larcher committed
403
                       100.0 * accuracy.item() / ((batch_idx + 1) * batch_size)))
404
405
406
    return model


Anthony Larcher's avatar
Anthony Larcher committed
407
def cross_validation(model, validation_loader, device):
408
409
410
411
412
413
414
415
416
417
418
419
420
    """

    :param args:
    :param model:
    :param cv_seg_df:
    :return:
    """
    model.eval()
    model.to(device)

    accuracy = 0.0
    criterion = torch.nn.CrossEntropyLoss()

Anthony Larcher's avatar
Anthony Larcher committed
421
    for batch_idx, (data, target) in enumerate(validation_loader):
Anthony Larcher's avatar
Anthony Larcher committed
422
        batch_size = target.shape[0]
423
424
425
426
427
428
        target = target.squeeze()
        output = model(data.to(device))
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()

    loss = criterion(output, target.to(device))

Anthony Larcher's avatar
Anthony Larcher committed
429
    return 100. * accuracy.cpu().numpy() / ((batch_idx + 1) * batch_size), loss
430
431
432


def xtrain_asynchronous(args):
433
434
435
436
437
438
    """
    Initialize and train an x-vector in asynchronous manner

    :param args:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
439
    # Initialize a first model and save to disk
Anthony Larcher's avatar
Anthony Larcher committed
440
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
441
442
443
444
    current_model_file_name = "initial_model"
    torch.save(model.state_dict(), current_model_file_name)

    for epoch in range(1, args.epochs + 1):
445
        current_model_file_name = train_asynchronous_epoch(epoch, args, current_model_file_name)
Anthony Larcher's avatar
Anthony Larcher committed
446
447

        # Add the cross validation here
448
        accuracy = cross_asynchronous_validation(args, current_model_file_name)
Anthony Larcher's avatar
Anthony Larcher committed
449
        print("*** Cross validation accuracy = {} %".format(accuracy))
Anthony Larcher's avatar
Anthony Larcher committed
450

Anthony Larcher's avatar
Anthony Larcher committed
451
        # Decrease learning rate after every epoch
Anthony Larcher's avatar
sad    
Anthony Larcher committed
452
453
        args.lr = args.lr * 0.9
        print("        Decrease learning rate: {}".format(args.lr))
Anthony Larcher's avatar
Anthony Larcher committed
454

Anthony Larcher's avatar
Anthony Larcher committed
455

456
def train_asynchronous_epoch(epoch, args, initial_model_file_name):
457
458
459
460
461
462
463
464
    """
    Process one training epoch using an asynchronous implementation of the training

    :param epoch:
    :param args:
    :param initial_model_file_name:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
    # Compute the megabatch number
    with open(args.batch_training_list, 'r') as fh:
        batch_file_list = [l.rstrip() for l in fh]

    # Shorten the batch_file_list to be a multiple of

    megabatch_number = len(batch_file_list) // (args.averaging_step * args.num_processes)
    megabatch_size = args.averaging_step * args.num_processes
    print("Epoch {}, number of megabatches = {}".format(epoch, megabatch_number))

    current_model = initial_model_file_name

    # For each sublist: run an asynchronous training and averaging of the model
    for ii in range(megabatch_number):
        print('Process megabatch [{} / {}]'.format(ii + 1, megabatch_number))
        current_model = train_asynchronous(epoch,
                                           args,
                                           current_model,
                                           batch_file_list[megabatch_size * ii: megabatch_size * (ii + 1)],
                                           ii,
485
                                           megabatch_number)  # function that split train, fuse and write the new model
Anthony Larcher's avatar
Anthony Larcher committed
486
487
488
    return current_model


489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
def train_asynchronous(epoch, args, initial_model_file_name, batch_file_list, megabatch_idx, megabatch_number):
    """
    Process one mega-batch of data asynchronously, average the model parameters across
    subrocesses and return the updated version of the model

    :param epoch:
    :param args:
    :param initial_model_file_name:
    :param batch_file_list:
    :param megabatch_idx:
    :param megabatch_number:
    :return:
    """
    # Split the list of files for each process
    sub_lists = split_file_list(batch_file_list, args.num_processes)

    #
    output_queue = mp.Queue()
    # output_queue = multiprocessing.Queue()

    processes = []
    for rank in range(args.num_processes):
511
        p = mp.Process(target=train_asynchronous_worker,
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
                       args=(rank, epoch, args, initial_model_file_name, sub_lists[rank], output_queue)
                       )
        # We first train the model across `num_processes` processes
        p.start()
        processes.append(p)

    # Average the models and write the new one to disk
    asynchronous_model = []
    for ii in range(args.num_processes):
        asynchronous_model.append(dict(output_queue.get()))

    for p in processes:
        p.join()

    av_model = Xtractor(args.class_number, args.dropout)
    tmp = av_model.state_dict()

    average_param = dict()
    for k in list(asynchronous_model[0].keys()):
        average_param[k] = asynchronous_model[0][k]

        for mod in asynchronous_model[1:]:
            average_param[k] += mod[k]

        if 'num_batches_tracked' not in k:
            tmp[k] = torch.FloatTensor(average_param[k] / len(asynchronous_model))

    # return the file name of the new model
    current_model_file_name = "{}/model_{}_epoch_{}_batch_{}".format(args.model_path, args.expe_id, epoch,
                                                                     megabatch_idx)
    torch.save(tmp, current_model_file_name)
    if megabatch_idx == megabatch_number:
        torch.save(tmp, "{}/model_{}_epoch_{}".format(args.model_path, args.expe_id, epoch))

    return current_model_file_name


549
def train_asynchronous_worker(rank, epoch, args, initial_model_file_name, batch_list, output_queue):
550
551
552
553
554
555
556
557
558
559
560
    """


    :param rank:
    :param epoch:
    :param args:
    :param initial_model_file_name:
    :param batch_list:
    :param output_queue:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
561
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
562
563
564
565
    model.load_state_dict(torch.load(initial_model_file_name))
    model.train()

    torch.manual_seed(args.seed + rank)
Anthony Larcher's avatar
Anthony Larcher committed
566
    train_loader = XvectorMultiDataset(batch_list, args.batch_path)
Anthony Larcher's avatar
Anthony Larcher committed
567
568
569
570
571
572
573
574
575
576
577
578

    device = torch.device("cuda:{}".format(rank))
    model.to(device)

    optimizer = optim.Adam([{'params': model.frame_conv0.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.frame_conv1.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.frame_conv2.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.frame_conv3.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.frame_conv4.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.seg_lin0.parameters(), 'weight_decay': args.l2_seg},
                            {'params': model.seg_lin1.parameters(), 'weight_decay': args.l2_seg},
                            {'params': model.seg_lin2.parameters(), 'weight_decay': args.l2_seg}
Anthony Larcher's avatar
Anthony Larcher committed
579
                            ], lr=args.lr)
Anthony Larcher's avatar
Anthony Larcher committed
580

Anthony Larcher's avatar
Anthony Larcher committed
581
    criterion = torch.nn.CrossEntropyLoss()
Anthony Larcher's avatar
Anthony Larcher committed
582
583
584
585
586
587
588
589

    accuracy = 0.0
    for batch_idx, (data, target) in enumerate(train_loader):
        optimizer.zero_grad()
        output = model(data.to(device))
        loss = criterion(output, target.to(device))
        loss.backward()
        optimizer.step()
Anthony Larcher's avatar
Anthony Larcher committed
590

Anthony Larcher's avatar
Anthony Larcher committed
591
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()
Anthony Larcher's avatar
Anthony Larcher committed
592

Anthony Larcher's avatar
Anthony Larcher committed
593
594
595
        if batch_idx % args.log_interval == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
                epoch, batch_idx + 1, train_loader.__len__(),
596
597
                100. * batch_idx / train_loader.__len__(), loss.item(),
                100.0 * accuracy.item() / ((batch_idx + 1) * args.batch_size)))
Anthony Larcher's avatar
Anthony Larcher committed
598

Anthony Larcher's avatar
Anthony Larcher committed
599
600
    model_param = OrderedDict()
    params = model.state_dict()
Anthony Larcher's avatar
Anthony Larcher committed
601

Anthony Larcher's avatar
Anthony Larcher committed
602
603
604
    for k in list(params.keys()):
        model_param[k] = params[k].cpu().detach().numpy()
    output_queue.put(model_param)
Anthony Larcher's avatar
Anthony Larcher committed
605
606


607
def cross_asynchronous_validation(args, current_model_file_name):
Anthony Larcher's avatar
Anthony Larcher committed
608
609
    """

Anthony Larcher's avatar
Anthony Larcher committed
610
611
    :param args:
    :param current_model_file_name:
Anthony Larcher's avatar
Anthony Larcher committed
612
613
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
614
    with open(args.cross_validation_list, 'r') as fh:
Anthony Larcher's avatar
Anthony Larcher committed
615
        cross_validation_list = [l.rstrip() for l in fh]
Anthony Larcher's avatar
Anthony Larcher committed
616
        sub_lists = split_file_list(cross_validation_list, args.num_processes)
Anthony Larcher's avatar
Anthony Larcher committed
617

Anthony Larcher's avatar
Anthony Larcher committed
618
619
    #
    output_queue = mp.Queue()
Anthony Larcher's avatar
Anthony Larcher committed
620

Anthony Larcher's avatar
Anthony Larcher committed
621
622
    processes = []
    for rank in range(args.num_processes):
623
        p = mp.Process(target=cv_asynchronous_worker,
Anthony Larcher's avatar
Anthony Larcher committed
624
625
626
627
628
                       args=(rank, args, current_model_file_name, sub_lists[rank], output_queue)
                       )
        # We first evaluate the model across `num_processes` processes
        p.start()
        processes.append(p)
Anthony Larcher's avatar
Anthony Larcher committed
629

Anthony Larcher's avatar
Anthony Larcher committed
630
631
632
633
    # Average the models and write the new one to disk
    result = []
    for ii in range(args.num_processes):
        result.append(output_queue.get())
Anthony Larcher's avatar
Anthony Larcher committed
634

Anthony Larcher's avatar
Anthony Larcher committed
635
636
    for p in processes:
        p.join()
Anthony Larcher's avatar
Anthony Larcher committed
637

Anthony Larcher's avatar
Anthony Larcher committed
638
639
640
    # Compute the global accuracy
    accuracy = 0.0
    total_batch_number = 0
Anthony Larcher's avatar
Anthony Larcher committed
641
    for bn, acc in result:
Anthony Larcher's avatar
Anthony Larcher committed
642
        accuracy += acc
Anthony Larcher's avatar
Anthony Larcher committed
643
644
        total_batch_number += bn
    
Anthony Larcher's avatar
Anthony Larcher committed
645
    return 100. * accuracy / (total_batch_number * args.batch_size)
Anthony Larcher's avatar
Anthony Larcher committed
646
647


648
def cv_asynchronous_worker(rank, args, current_model_file_name, batch_list, output_queue):
Anthony Larcher's avatar
Anthony Larcher committed
649
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
650
651
    model.load_state_dict(torch.load(current_model_file_name))
    model.eval()
Anthony Larcher's avatar
Anthony Larcher committed
652

Anthony Larcher's avatar
Anthony Larcher committed
653
    cv_loader = XvectorMultiDataset(batch_list, args.batch_path)
Anthony Larcher's avatar
Anthony Larcher committed
654

Anthony Larcher's avatar
Anthony Larcher committed
655
656
    device = torch.device("cuda:{}".format(rank))
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
657

Anthony Larcher's avatar
Anthony Larcher committed
658
    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
659
    for batch_idx, (data, target) in enumerate(cv_loader):
Anthony Larcher's avatar
Anthony Larcher committed
660
661
        output = model(data.to(device))
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()
Anthony Larcher's avatar
Anthony Larcher committed
662
    output_queue.put((cv_loader.__len__(), accuracy.cpu().numpy()))
Anthony Larcher's avatar
Anthony Larcher committed
663

Anthony Larcher's avatar
hot    
Anthony Larcher committed
664

665
def extract_idmap(args, device_id, segment_indices, fs_params, idmap_name, output_queue):
Anthony Larcher's avatar
Anthony Larcher committed
666
    """
Anthony Larcher's avatar
Anthony Larcher committed
667
668
    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
669
    """
670
    # device = torch.device("cuda:{}".format(device_ID))
Anthony Larcher's avatar
Anthony Larcher committed
671
    device = torch.device('cpu')
Anthony Larcher's avatar
Anthony Larcher committed
672
673
674
675
676
677
678
679
680
681
682
683
684

    # 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
685
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
686
687
688
689
690
691
692
693
    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
694
695
696
697
698
699
    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
700
701
702
703
704
705

    # 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
706
        logging.critical('Process file {}, [{} / {}]'.format(segment_id, idx, segment_loader.__len__()))
Anthony Larcher's avatar
Anthony Larcher committed
707

Anthony Larcher's avatar
Anthony Larcher committed
708
709
710
        if list(data.shape)[2] < 20:
            pass
        else:
Anthony Larcher's avatar
Anthony Larcher committed
711
712
713
714
715
716
717
            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
718

Anthony Larcher's avatar
Anthony Larcher committed
719
    output_queue.put((segment_indices, emb_1, emb_2, emb_3, emb_4, emb_5, emb_6))
Anthony Larcher's avatar
Anthony Larcher committed
720
721


Anthony Larcher's avatar
Anthony Larcher committed
722
def extract_parallel(args, fs_params):
723
724
725
726
727
728
    """

    :param args:
    :param fs_params:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
729
730
731
    emb_a_size = 512
    emb_b_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
732
    idmap = IdMap(args.idmap)
Anthony Larcher's avatar
Anthony Larcher committed
733

Anthony Larcher's avatar
Anthony Larcher committed
734
735
736
737
738
739
740
741
742
743
744
745
746
    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
747
748
749

    # Split the indices
    mega_batch_size = idmap.leftids.shape[0] // args.num_processes
Anthony Larcher's avatar
Anthony Larcher committed
750
751
752

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

Anthony Larcher's avatar
Anthony Larcher committed
753
754
755
    segment_idx = []
    for ii in range(args.num_processes):
        segment_idx.append(
Anthony Larcher's avatar
Anthony Larcher committed
756
757
758
759
            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
760
761
762
763
764
765
766

    # 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
767
                       args=(args, rank, segment_idx[rank], fs_params, args.idmap, output_queue)
Anthony Larcher's avatar
Anthony Larcher committed
768
769
770
771
772
773
774
                       )
        # 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
775
776
777
778
779
780
781
        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
782
783
784
785

    for p in processes:
        p.join()

Anthony Larcher's avatar
Anthony Larcher committed
786
    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
787
788


Anthony Larcher's avatar
Anthony Larcher committed
789
def extract_embeddings(args):
790
791
792
793
794
795
796
797
798
799
800
801
802
    """

    :param args:
    :param device_id:
    :param fs_params:
    :return:
    """
    device = torch.device("cuda:0")

    # Load the model
    logging.critical("*** Load model from = {}/{}".format(args.model_path, args.init_model_name))
    model_file_name = '/'.join([args.model_path, args.init_model_name])
    model = torch.load(model_file_name)
Anthony Larcher's avatar
Anthony Larcher committed
803
    model = torch.nn.DataParallel(model)
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
    model.eval()
    model.to(device)

    # Get the list of files
    total_seg_df = pickle.load(open(args.batch_training_list, "rb"))

    speaker_dict = {}
    tmp = total_seg_df.speaker_id.unique()
    tmp.sort()
    for idx, spk in enumerate(tmp):
        speaker_dict[spk] = idx

    extract_transform = [CMVN(), ]
    extract_set = VoxDataset(total_seg_df, speaker_dict, None, transform=transforms.Compose(extract_transform),
                             spec_aug_ratio=args.spec_aug, temp_aug_ratio=args.temp_aug)
    extract_loader = DataLoader(extract_set, batch_size=1, shuffle=False, num_workers=5)
Anthony Larcher's avatar
Anthony Larcher committed
820

821
    #CREER UN TENSEUR DE LA BONNE TAILLE POUR STOCKER LES X-VECTEURS
Anthony Larcher's avatar
Anthony Larcher committed
822

823
824
825
826
    for batch_idx, (data, target, _, __) in enumerate(extract_loader):
        print("extrait x-vecteur numero {}".format(batch_idx))
        embedding = model.produce_embeddings(data.to(device))
        #REMPLIR LE TENSEUR AVEC LE NOUVEAU X-VECTEUR
Anthony Larcher's avatar
Anthony Larcher committed
827

828
829
    #FAIRE CORRESPONDRE LES SPK_ID AVEC LES X-VECTEURS
    #RENVOYER LE TENSEUR DE X-VECTEURS SUR LE CPU OU L ECRTIRE SUR LE DISQUE