xvector.py 32.9 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
minor    
Anthony Larcher committed
30
import pickle
Anthony Larcher's avatar
Anthony Larcher committed
31
import torch
Anthony Larcher's avatar
Anthony Larcher committed
32
33
import torch.optim as optim
import torch.multiprocessing as mp
Anthony Larcher's avatar
Anthony Larcher committed
34
from torchvision import transforms
Anthony Larcher's avatar
Anthony Larcher committed
35
from collections import OrderedDict
36
from .xsets import XvectorMultiDataset, StatDataset, VoxDataset
Anthony Larcher's avatar
Anthony Larcher committed
37
38
39
from .xsets import FrequencyMask, CMVN, TemporalMask
from ..bosaris import IdMap
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
40
from torch.utils.data import DataLoader
Anthony Larcher's avatar
Anthony Larcher committed
41

Anthony Larcher's avatar
Anthony Larcher committed
42
43
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
44
__copyright__ = "Copyright 2015-2020 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
45
46
47
48
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
49
50


Anthony Larcher's avatar
Anthony Larcher committed
51
# logging.basicConfig(stream=sys.stdout, level=logging.INFO)
Anthony Larcher's avatar
Anthony Larcher committed
52
53


54
55
56
57
58
def get_lr(optimizer):
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
59
60
61
62
63
64
65
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
66
67
68


class Xtractor(torch.nn.Module):
69
70
71
    """
    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
72

Anthony Larcher's avatar
Anthony Larcher committed
73
74
75
76
77
78
    def __init__(self, speaker_number, config=None):
        """
        If config is None, default architecture is created
        :param config:
        """
        self.speaker_number = speaker_number
Anthony Larcher's avatar
xv    
Anthony Larcher committed
79
        self.activation = torch.nn.ReLU()
Anthony Larcher's avatar
Anthony Larcher committed
80
        if config is None:
Anthony Larcher's avatar
xv    
Anthony Larcher committed
81
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
                ("conv1", torch.nn.Conv1d(30, 512, 5, dilation=1)),
                ("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)),
                ("conv4", torch.nn.Conv1d(512, 512)),
                ("activation4", torch.nn.LeakyReLU(0.2)),
                ("norm4", torch.nn.BatchNorm1d(512)),
                ("conv5", torch.nn.Conv1d(512, 1536)),
                ("activation5", torch.nn.LeakyReLU(0.2)),
                ("norm5", torch.nn.BatchNorm1d(1536))
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
99
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
100
101
102
                ("linear6", torch.nn.linear(1536, 512))
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
103
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
                ("activation6", torch.nn.LeakyReLU(0.2)),
                ("norm6", torch.nn.BatchNorm1d(512)),
                ("linear7", torch.nn.linear(512, 512)),
                ("activation7", torch.nn.LeakyReLU(0.2)),
                ("norm7", torch.nn.BatchNorm1d(512)),
                ("linear8", torch.nn.linear(512, self.speaker_number ))
            ]))

        else:
            # Load Yaml configuration
            with open(config, 'r') as fh:
                cfg = yaml.load(fh, Loader=yaml.FullLoader)

            # Get Feature size
            self.feature_size = cfg["feature_size"]
            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"):
                    segmental_layers.append((k, torch.nn.Conv2d(input_size,
                                                                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)))

            self.sequence_network = nn.Sequential(OrderedDict(segmental_layers))

            # Create sequential object for the second part of the network
            input_size = input_size * 2
Anthony Larcher's avatar
xv    
Anthony Larcher committed
151
152
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
153
154
                if k.startswith("lin"):
                    if cfg["embedding"][k]["output"] == "speaker_number":
Anthony Larcher's avatar
xv    
Anthony Larcher committed
155
                        before_embedding_layers.append((k, torch.nn.linear(input_size, self.speaker_number)))
Anthony Larcher's avatar
Anthony Larcher committed
156
                    else:
Anthony Larcher's avatar
xv    
Anthony Larcher committed
157
                        before_embedding_layers.append((k, torch.nn.linear(input_size, cfg["embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
158
159
160
                        input_size = cfg["embedding"][k]["output"]

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
161
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
162
163

                elif k.startswith('norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
164
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
165
166

                elif k.startswith('dropout'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["emebedding"][k])))

            self.before_speaker_embedding = nn.Sequential(OrderedDict(before_embedding_layers))

            # Create sequential object for the second part of the network
            after_embedding_layers = []
            for k in cfg["after_embedding"].keys():
                if k.startswith("lin"):
                    if cfg["embedding"][k]["output"] == "speaker_number":
                        after_embedding_layers.append((k, torch.nn.linear(input_size, self.speaker_number)))
                    else:
                        after_embedding_layers.append((k, torch.nn.linear(input_size, cfg["embedding"][k]["output"])))
                        input_size = cfg["embedding"][k]["output"]

                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'):
                    after_embedding_layers.append((k, torch.nn.Dropout(p=cfg["emebedding"][k])))

            self.after_embedding_layers = nn.Sequential(OrderedDict(after_embedding_layers))

Anthony Larcher's avatar
Anthony Larcher committed
192

Anthony Larcher's avatar
Anthony Larcher committed
193

194
    def produce_embeddings(self, x):
Anthony Larcher's avatar
Anthony Larcher committed
195
        """
Anthony Larcher's avatar
Anthony Larcher committed
196

197
198
199
        :param x:
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
200
201
202
203
204
        frame_emb_0 = self.norm0(self.activation(self.frame_conv0(x)))
        frame_emb_1 = self.norm1(self.activation(self.frame_conv1(frame_emb_0)))
        frame_emb_2 = self.norm2(self.activation(self.frame_conv2(frame_emb_1)))
        frame_emb_3 = self.norm3(self.activation(self.frame_conv3(frame_emb_2)))
        frame_emb_4 = self.norm4(self.activation(self.frame_conv4(frame_emb_3)))
Anthony Larcher's avatar
Anthony Larcher committed
205
206
207

        mean = torch.mean(frame_emb_4, dim=2)
        std = torch.std(frame_emb_4, dim=2)
208
        seg_emb = torch.cat([mean, std], dim=1)
Anthony Larcher's avatar
Anthony Larcher committed
209

210
211
        embedding_a = self.seg_lin0(seg_emb)
        return embedding_a
Anthony Larcher's avatar
Anthony Larcher committed
212

Anthony Larcher's avatar
Anthony Larcher committed
213
    def forward(self, x, is_eval=False):
214
215
216
217
        """

        :param x:
        :return:
Anthony Larcher's avatar
Anthony Larcher committed
218

219
        seg_emb_0 = self.produce_embeddings(x)
Anthony Larcher's avatar
Anthony Larcher committed
220
        # batch-normalisation after this layer
221
        seg_emb_1 = self.norm6(self.activation(seg_emb_0))
Anthony Larcher's avatar
Anthony Larcher committed
222
        # new layer with batch Normalization
223
        seg_emb_2 = self.norm7(self.activation(self.seg_lin1(self.dropout_lin1(seg_emb_1))))
Anthony Larcher's avatar
Anthony Larcher committed
224
        # No batch-normalisation after this layer
Anthony Larcher's avatar
minor    
Anthony Larcher committed
225
        result = self.seg_lin2(seg_emb_2)
Anthony Larcher's avatar
Anthony Larcher committed
226
        return result
227
        """
Anthony Larcher's avatar
Anthony Larcher committed
228
        x = self.sequence_network(x)
229

Anthony Larcher's avatar
Anthony Larcher committed
230
231
232
233
234
235
236
237
        # 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
238

Anthony Larcher's avatar
Anthony Larcher committed
239
240
        x = self.after_speaker_embedding(x)
        return x
Anthony Larcher's avatar
Anthony Larcher committed
241
242
243

    def init_weights(self):
        """
244
        Initialize the x-vector extract weights and biaises
Anthony Larcher's avatar
Anthony Larcher committed
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
        """
        torch.nn.init.normal_(self.frame_conv0.weight, mean=-0.5, std=0.1)
        torch.nn.init.normal_(self.frame_conv1.weight, mean=-0.5, std=0.1)
        torch.nn.init.normal_(self.frame_conv2.weight, mean=-0.5, std=0.1)
        torch.nn.init.normal_(self.frame_conv3.weight, mean=-0.5, std=0.1)
        torch.nn.init.normal_(self.frame_conv4.weight, mean=-0.5, std=0.1)
        torch.nn.init.xavier_uniform(self.seg_lin0.weight)
        torch.nn.init.xavier_uniform(self.seg_lin1.weight)
        torch.nn.init.xavier_uniform(self.seg_lin2.weight)

        torch.nn.init.constant(self.frame_conv0.bias, 0.1)
        torch.nn.init.constant(self.frame_conv1.bias, 0.1)
        torch.nn.init.constant(self.frame_conv2.bias, 0.1)
        torch.nn.init.constant(self.frame_conv3.bias, 0.1)
        torch.nn.init.constant(self.frame_conv4.bias, 0.1)
        torch.nn.init.constant(self.seg_lin0.bias, 0.1)
        torch.nn.init.constant(self.seg_lin1.bias, 0.1)
        torch.nn.init.constant(self.seg_lin2.bias, 0.1)

Anthony Larcher's avatar
Anthony Larcher committed
264
265

def xtrain(args):
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
    """
    Initialize and train an x-vector on a single GPU

    :param args:
    :return:
    """
    # If we start from an existing model
    if not args.init_model_name == '':
        # 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)
        model.train()
    else:
        # Initialize a first model and save to disk
        model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
minor    
Anthony Larcher committed
282
        model.init_weights()
283
        model.train()
Anthony Larcher's avatar
Anthony Larcher committed
284
285
286
287
288

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

289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
    model.cuda()

    # Split the training data in train and cv
    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
    pickle.dump(speaker_dict, open("spk_dictionary.pkl", "wb"))

    cv_portion = 0.007
    idx = numpy.arange(len(total_seg_df))
    numpy.random.shuffle(idx)
    train_seg_df = total_seg_df.iloc[idx[:int((1 - cv_portion) * len(idx))]].reset_index()
    cv_seg_df = total_seg_df.iloc[idx[int((1 - cv_portion) * len(idx)):]].reset_index()

    current_model_file_name = "initial_model"
    torch.save(model.state_dict(), current_model_file_name)

Anthony Larcher's avatar
minor    
Anthony Larcher committed
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
    if type(model) is Xtractor:
        optimizer = torch.optim.SGD([
            {'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}
            ],
            lr=args.lr, momentum=0.9)
    else:
        optimizer = torch.optim.SGD([
            {'params': model.module.frame_conv0.parameters(), 'weight_decay': args.l2_frame},
            {'params': model.module.frame_conv1.parameters(), 'weight_decay': args.l2_frame},
            {'params': model.module.frame_conv2.parameters(), 'weight_decay': args.l2_frame},
            {'params': model.module.frame_conv3.parameters(), 'weight_decay': args.l2_frame},
            {'params': model.module.frame_conv4.parameters(), 'weight_decay': args.l2_frame},
            {'params': model.module.seg_lin0.parameters(), 'weight_decay': args.l2_seg},
            {'params': model.module.seg_lin1.parameters(), 'weight_decay': args.l2_seg},
            {'params': model.module.seg_lin2.parameters(), 'weight_decay': args.l2_seg}
            ],
            lr=args.lr, momentum=0.9)
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383

    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min')

    for epoch in range(1, args.epochs + 1):
        # Process one epoch and return the current model
        model = train_epoch(model, epoch, train_seg_df, speaker_dict, optimizer, args)

        # Add the cross validation here
        accuracy, val_loss = cross_validation(args, model, cv_seg_df, speaker_dict)
        logging.critical("*** Cross validation accuracy = {} %".format(accuracy))

        # Decrease learning rate according to the scheduler policy
        scheduler.step(val_loss)

        # return the file name of the new model
        base_name = "model"
        if not args.init_model_name == "":
            base_name = args.init_model_name
        current_model_file_name = "{}/{}_{}_epoch_{}".format(args.model_path, base_name, args.expe_id, epoch)
        torch.save(model, current_model_file_name)


def train_epoch(model, epoch, train_seg_df, speaker_dict, optimizer, args):
    """

    :param model:
    :param epoch:
    :param train_seg_df:
    :param speaker_dict:
    :param optimizer:
    :param args:
    :return:
    """
    device = torch.device("cuda:0")

    torch.manual_seed(args.seed)

    train_transform = []
    if not args.train_transformation == '':
        trans = args.train_transformation.split(',')
        for t in trans:
            if "CMVN" in t:
                train_transform.append(CMVN())
            if "FrequencyMask" in t:
                a = int(t.split("-")[0].split("(")[1])
                b = int(t.split("-")[1].split(")")[0])
                train_transform.append(FrequencyMask(a, b))
            if "TemporalMask" in t:
                a = int(t.split("(")[1].split(")")[0])
                train_transform.append(TemporalMask(a))
Anthony Larcher's avatar
Anthony Larcher committed
384
    train_set = VoxDataset(train_seg_df, speaker_dict, args.duration, transform=transforms.Compose(train_transform),
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
                           spec_aug_ratio=args.spec_aug, temp_aug_ratio=args.temp_aug)
    train_loader = DataLoader(train_set, batch_size=args.batch_size, shuffle=True, num_workers=15)

    criterion = torch.nn.CrossEntropyLoss()

    accuracy = 0.0
    for batch_idx, (data, target, _, __) in enumerate(train_loader):
        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()

        if batch_idx % args.log_interval == 0:
            logging.critical('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
                epoch, batch_idx + 1, train_loader.__len__(),
                       100. * batch_idx / train_loader.__len__(), loss.item(),
                       100.0 * accuracy.item() / ((batch_idx + 1) * args.batch_size)))
    return model


# def cross_validation(args, model):
#
#     with open(args.cross_validation_list, 'r') as fh:
#         cross_validation_list = [l.rstrip() for l in fh]
#     cv_loader = XvectorMultiDataset(cross_validation_list, args.batch_path)
#
#     model.eval()
#     device = torch.device("cuda:0")
#     model.to(device)
#
#     accuracy = 0.0
#     bi = 0
#     for batch_idx, (data, target) in enumerate(cv_loader):
#         output = model(data.to(device))
#         accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()
#         bi = batch_idx
#     return 100. * accuracy.cpu().numpy() / ((bi + 1) * args.batch_size)


def cross_validation(args, model, cv_seg_df, speaker_dict):
    """

    :param args:
    :param model:
    :param cv_seg_df:
    :return:
    """
    cv_transform = []
    if not args.cv_transformation == '':
        trans = args.cv_transformation.split(',')
        for t in trans:
            if "CMVN" in t:
                cv_transform.append(CMVN())
            if "FrequencyMask" in t:
                a = t.split(",")[0].split("(")[1]
                b = t.split(",")[1].split("(")[0]
                cv_transform.append(FrequencyMask(a, b))
            if "TemporalMask" in t:
                a = t.split(",")[0].split("(")[1]
                cv_transform.append(TemporalMask(a, b))
    cv_set = VoxDataset(cv_seg_df, speaker_dict, 500, transform=transforms.Compose(cv_transform),
                        spec_aug_ratio=args.spec_aug, temp_aug_ratio=args.temp_aug)
    cv_loader = DataLoader(cv_set, batch_size=args.batch_size, shuffle=False, num_workers=15)
    model.eval()
    device = torch.device("cuda:0")
    model.to(device)

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

    for batch_idx, (data, target, _, __) in enumerate(cv_loader):
        target = target.squeeze()
        output = model(data.to(device))
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()

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

    return 100. * accuracy.cpu().numpy() / ((batch_idx + 1) * args.batch_size), loss


def xtrain_asynchronous(args):
469
470
471
472
473
474
    """
    Initialize and train an x-vector in asynchronous manner

    :param args:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
475
    # Initialize a first model and save to disk
Anthony Larcher's avatar
Anthony Larcher committed
476
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
477
478
479
480
    current_model_file_name = "initial_model"
    torch.save(model.state_dict(), current_model_file_name)

    for epoch in range(1, args.epochs + 1):
481
        current_model_file_name = train_asynchronous_epoch(epoch, args, current_model_file_name)
Anthony Larcher's avatar
Anthony Larcher committed
482
483

        # Add the cross validation here
484
        accuracy = cross_asynchronous_validation(args, current_model_file_name)
Anthony Larcher's avatar
Anthony Larcher committed
485
        print("*** Cross validation accuracy = {} %".format(accuracy))
Anthony Larcher's avatar
Anthony Larcher committed
486

Anthony Larcher's avatar
Anthony Larcher committed
487
        # Decrease learning rate after every epoch
Anthony Larcher's avatar
sad    
Anthony Larcher committed
488
489
        args.lr = args.lr * 0.9
        print("        Decrease learning rate: {}".format(args.lr))
Anthony Larcher's avatar
Anthony Larcher committed
490

Anthony Larcher's avatar
Anthony Larcher committed
491

492
def train_asynchronous_epoch(epoch, args, initial_model_file_name):
493
494
495
496
497
498
499
500
    """
    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
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
    # 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,
521
                                           megabatch_number)  # function that split train, fuse and write the new model
Anthony Larcher's avatar
Anthony Larcher committed
522
523
524
    return current_model


525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
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):
547
        p = mp.Process(target=train_asynchronous_worker,
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
                       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


585
def train_asynchronous_worker(rank, epoch, args, initial_model_file_name, batch_list, output_queue):
586
587
588
589
590
591
592
593
594
595
596
    """


    :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
597
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
598
599
600
601
    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
602
    train_loader = XvectorMultiDataset(batch_list, args.batch_path)
Anthony Larcher's avatar
Anthony Larcher committed
603
604
605
606
607
608
609
610
611
612
613
614

    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
615
                            ], lr=args.lr)
Anthony Larcher's avatar
Anthony Larcher committed
616

Anthony Larcher's avatar
Anthony Larcher committed
617
    criterion = torch.nn.CrossEntropyLoss()
Anthony Larcher's avatar
Anthony Larcher committed
618
619
620
621
622
623
624
625

    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
626

Anthony Larcher's avatar
Anthony Larcher committed
627
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()
Anthony Larcher's avatar
Anthony Larcher committed
628

Anthony Larcher's avatar
Anthony Larcher committed
629
630
631
        if batch_idx % args.log_interval == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
                epoch, batch_idx + 1, train_loader.__len__(),
632
633
                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
634

Anthony Larcher's avatar
Anthony Larcher committed
635
636
    model_param = OrderedDict()
    params = model.state_dict()
Anthony Larcher's avatar
Anthony Larcher committed
637

Anthony Larcher's avatar
Anthony Larcher committed
638
639
640
    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
641
642


643
def cross_asynchronous_validation(args, current_model_file_name):
Anthony Larcher's avatar
Anthony Larcher committed
644
645
    """

Anthony Larcher's avatar
Anthony Larcher committed
646
647
    :param args:
    :param current_model_file_name:
Anthony Larcher's avatar
Anthony Larcher committed
648
649
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
650
    with open(args.cross_validation_list, 'r') as fh:
Anthony Larcher's avatar
Anthony Larcher committed
651
        cross_validation_list = [l.rstrip() for l in fh]
Anthony Larcher's avatar
Anthony Larcher committed
652
        sub_lists = split_file_list(cross_validation_list, args.num_processes)
Anthony Larcher's avatar
Anthony Larcher committed
653

Anthony Larcher's avatar
Anthony Larcher committed
654
655
    #
    output_queue = mp.Queue()
Anthony Larcher's avatar
Anthony Larcher committed
656

Anthony Larcher's avatar
Anthony Larcher committed
657
658
    processes = []
    for rank in range(args.num_processes):
659
        p = mp.Process(target=cv_asynchronous_worker,
Anthony Larcher's avatar
Anthony Larcher committed
660
661
662
663
664
                       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
665

Anthony Larcher's avatar
Anthony Larcher committed
666
667
668
669
    # 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
670

Anthony Larcher's avatar
Anthony Larcher committed
671
672
    for p in processes:
        p.join()
Anthony Larcher's avatar
Anthony Larcher committed
673

Anthony Larcher's avatar
Anthony Larcher committed
674
675
676
    # Compute the global accuracy
    accuracy = 0.0
    total_batch_number = 0
Anthony Larcher's avatar
Anthony Larcher committed
677
    for bn, acc in result:
Anthony Larcher's avatar
Anthony Larcher committed
678
        accuracy += acc
Anthony Larcher's avatar
Anthony Larcher committed
679
680
        total_batch_number += bn
    
Anthony Larcher's avatar
Anthony Larcher committed
681
    return 100. * accuracy / (total_batch_number * args.batch_size)
Anthony Larcher's avatar
Anthony Larcher committed
682
683


684
def cv_asynchronous_worker(rank, args, current_model_file_name, batch_list, output_queue):
Anthony Larcher's avatar
Anthony Larcher committed
685
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
686
687
    model.load_state_dict(torch.load(current_model_file_name))
    model.eval()
Anthony Larcher's avatar
Anthony Larcher committed
688

Anthony Larcher's avatar
Anthony Larcher committed
689
    cv_loader = XvectorMultiDataset(batch_list, args.batch_path)
Anthony Larcher's avatar
Anthony Larcher committed
690

Anthony Larcher's avatar
Anthony Larcher committed
691
692
    device = torch.device("cuda:{}".format(rank))
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
693

Anthony Larcher's avatar
Anthony Larcher committed
694
    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
695
    for batch_idx, (data, target) in enumerate(cv_loader):
Anthony Larcher's avatar
Anthony Larcher committed
696
697
        output = model(data.to(device))
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()
Anthony Larcher's avatar
Anthony Larcher committed
698
    output_queue.put((cv_loader.__len__(), accuracy.cpu().numpy()))
Anthony Larcher's avatar
Anthony Larcher committed
699

Anthony Larcher's avatar
hot    
Anthony Larcher committed
700

701
def extract_idmap(args, device_id, segment_indices, fs_params, idmap_name, output_queue):
Anthony Larcher's avatar
Anthony Larcher committed
702
    """
Anthony Larcher's avatar
Anthony Larcher committed
703
704
    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
705
    """
706
    # device = torch.device("cuda:{}".format(device_ID))
Anthony Larcher's avatar
Anthony Larcher committed
707
    device = torch.device('cpu')
Anthony Larcher's avatar
Anthony Larcher committed
708
709
710
711
712
713
714
715
716
717
718
719
720

    # 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
721
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
722
723
724
725
726
727
728
729
    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
730
731
732
733
734
735
    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
736
737
738
739
740
741

    # 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
742
        logging.critical('Process file {}, [{} / {}]'.format(segment_id, idx, segment_loader.__len__()))
Anthony Larcher's avatar
Anthony Larcher committed
743

Anthony Larcher's avatar
Anthony Larcher committed
744
745
746
        if list(data.shape)[2] < 20:
            pass
        else:
Anthony Larcher's avatar
Anthony Larcher committed
747
748
749
750
751
752
753
            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
754

Anthony Larcher's avatar
Anthony Larcher committed
755
    output_queue.put((segment_indices, emb_1, emb_2, emb_3, emb_4, emb_5, emb_6))
Anthony Larcher's avatar
Anthony Larcher committed
756
757


Anthony Larcher's avatar
Anthony Larcher committed
758
def extract_parallel(args, fs_params):
759
760
761
762
763
764
    """

    :param args:
    :param fs_params:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
765
766
767
    emb_a_size = 512
    emb_b_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
768
    idmap = IdMap(args.idmap)
Anthony Larcher's avatar
Anthony Larcher committed
769

Anthony Larcher's avatar
Anthony Larcher committed
770
771
772
773
774
775
776
777
778
779
780
781
782
    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
783
784
785

    # Split the indices
    mega_batch_size = idmap.leftids.shape[0] // args.num_processes
Anthony Larcher's avatar
Anthony Larcher committed
786
787
788

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

Anthony Larcher's avatar
Anthony Larcher committed
789
790
791
    segment_idx = []
    for ii in range(args.num_processes):
        segment_idx.append(
Anthony Larcher's avatar
Anthony Larcher committed
792
793
794
795
            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
796
797
798
799
800
801
802

    # 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
803
                       args=(args, rank, segment_idx[rank], fs_params, args.idmap, output_queue)
Anthony Larcher's avatar
Anthony Larcher committed
804
805
806
807
808
809
810
                       )
        # 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
811
812
813
814
815
816
817
        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
818
819
820
821

    for p in processes:
        p.join()

Anthony Larcher's avatar
Anthony Larcher committed
822
    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
823
824


Anthony Larcher's avatar
Anthony Larcher committed
825
def extract_embeddings(args):
826
827
828
829
830
831
832
833
834
835
836
837
838
    """

    :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
839
    model = torch.nn.DataParallel(model)
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
    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
856

857
    #CREER UN TENSEUR DE LA BONNE TAILLE POUR STOCKER LES X-VECTEURS
Anthony Larcher's avatar
Anthony Larcher committed
858

859
860
861
862
    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
863

864
865
    #FAIRE CORRESPONDRE LES SPK_ID AVEC LES X-VECTEURS
    #RENVOYER LE TENSEUR DE X-VECTEURS SUR LE CPU OU L ECRTIRE SUR LE DISQUE