xvector.py 22 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
from .sincnet import SincNet
Anthony Larcher's avatar
Anthony Larcher committed
47

Anthony Larcher's avatar
Anthony Larcher committed
48
49
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
50
__copyright__ = "Copyright 2015-2020 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
51
52
53
54
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
55
56


Anthony Larcher's avatar
Anthony Larcher committed
57
# logging.basicConfig(stream=sys.stdout, level=logging.INFO)
Anthony Larcher's avatar
Anthony Larcher committed
58
59


60
61
62
63
64
def get_lr(optimizer):
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
65
66
67
68
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
Anthony Larcher committed
69
70
71


class Xtractor(torch.nn.Module):
72
73
74
    """
    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
75

Anthony Larcher's avatar
Anthony Larcher committed
76
    def __init__(self, speaker_number, model_archi=None):
Anthony Larcher's avatar
Anthony Larcher committed
77
78
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
79
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
80
        """
Anthony Larcher's avatar
Anthony Larcher committed
81
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
82
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
83
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
84

Anthony Larcher's avatar
Anthony Larcher committed
85
        if model_archi is None:
Anthony Larcher's avatar
Anthony Larcher committed
86
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
87
88
89
            self.activation = torch.nn.LeakyReLU(0.2)

            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
90

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(3072, 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
                ("dropout6", torch.nn.Dropout(p=0.05)),
Anthony Larcher's avatar
Anthony Larcher committed
117
                ("linear7", torch.nn.Linear(512, 512)),
Anthony Larcher's avatar
Anthony Larcher committed
118
119
                ("activation7", torch.nn.LeakyReLU(0.2)),
                ("norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
120
                ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
Anthony Larcher's avatar
Anthony Larcher committed
121
122
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
123
124
125
126
            self.sequence_network_weight_decay = 0.0002
            self.before_speaker_embedding_weight_decay = 0.002
            self.after_speaker_embedding_weight_decay = 0.002

Anthony Larcher's avatar
Anthony Larcher committed
127
128
        else:
            # Load Yaml configuration
Anthony Larcher's avatar
Anthony Larcher committed
129
            with open(model_archi, 'r') as fh:
Anthony Larcher's avatar
Anthony Larcher committed
130
131
                cfg = yaml.load(fh, Loader=yaml.FullLoader)

Anthony Larcher's avatar
Anthony Larcher committed
132
133
134
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
135
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
136
137
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
138
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
139
140
141
142
143
144
145
146
147
148
149
150
                        waveform_normalize=cfg['preprocessor']["waveform_normalize"],
                        sample_rate=cfg['preprocessor']["sample_rate"],
                        min_low_hz=cfg['preprocessor']["min_low_hz"],
                        min_band_hz=cfg['preprocessor']["min_band_hz"],
                        out_channels=cfg['preprocessor']["out_channels"],
                        kernel_size=cfg['preprocessor']["kernel_size"],
                        stride=cfg['preprocessor']["stride"],
                        max_pool=cfg['preprocessor']["max_pool"],
                        instance_normalize=cfg['preprocessor']["instance_normalize"],
                        activation=cfg['preprocessor']["activation"],
                        dropout=cfg['preprocessor']["dropout"]
                    )
Anthony Larcher's avatar
Anthony Larcher committed
151
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
152
153

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
154
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
155
            """
Anthony Larcher's avatar
Anthony Larcher committed
156
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
157
158
159
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
            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
176
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
177
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
178
179
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
180
181
182
183
184
185
186
187
                    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
188
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
189
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
190

Anthony Larcher's avatar
Anthony Larcher committed
191
192
193
            """
            Prepapre last part of the network (after pooling)
            """
Anthony Larcher's avatar
Anthony Larcher committed
194
195
            # Create sequential object for the second part of the network
            input_size = input_size * 2
Anthony Larcher's avatar
xv    
Anthony Larcher committed
196
197
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
198
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
199
200
                    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
201
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
202
203
                        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
204
205

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
206
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
207
208

                elif k.startswith('norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
209
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
210
211

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
212
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
213

Anthony Larcher's avatar
Anthony Larcher committed
214
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
215
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
216
217
218
219
220

            # 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
221
222
                    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
223
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
224
225
                        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
226
227
228
229
230
231
232
233

                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
234
                    after_embedding_layers.append((k, torch.nn.Dropout(p=cfg["after_embedding"][k])))
Anthony Larcher's avatar
Anthony Larcher committed
235

Anthony Larcher's avatar
Anthony Larcher committed
236
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict(after_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
237
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
238

Anthony Larcher's avatar
Anthony Larcher committed
239
    def forward(self, x, is_eval=False):
240
241
242
243
244
        """

        :param x:
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
245
246
247
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
248
        x = self.sequence_network(x)
249

Anthony Larcher's avatar
Anthony Larcher committed
250
251
252
253
254
255
256
257
        # 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
258

Anthony Larcher's avatar
Anthony Larcher committed
259
260
        x = self.after_speaker_embedding(x)
        return x
Anthony Larcher's avatar
Anthony Larcher committed
261

Anthony Larcher's avatar
Anthony Larcher committed
262

Anthony Larcher's avatar
Anthony Larcher committed
263

Anthony Larcher's avatar
minor    
Anthony Larcher committed
264

Anthony Larcher's avatar
Anthony Larcher committed
265
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
266
           dataset_yaml,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
267
           epochs=100,
Anthony Larcher's avatar
Anthony Larcher committed
268
           lr=0.01,
Anthony Larcher's avatar
Anthony Larcher committed
269
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
270
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
271
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
272
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
273
           multi_gpu=True,
274
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
275
           num_thread=1):
276
277
278
279
280
281
    """
    Initialize and train an x-vector on a single GPU

    :param args:
    :return:
    """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
282
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
283

284
    # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
285
286
287
288
289
290
291
292
    #if model_name is not None:
    #    # Load the model
    #    logging.critical(f"*** Load model from = {model_name}")
    #    checkpoint = torch.load(model_name)
    #    model = Xtractor(speaker_number, model_yaml)
    #    model.load_state_dict(checkpoint["model_state_dict"])
    #else:
    if True:
Anthony Larcher's avatar
Anthony Larcher committed
293
294
        # Initialize a first model
        if model_yaml is None:
Anthony Larcher's avatar
Anthony Larcher committed
295
            model = Xtractor(speaker_number)
Anthony Larcher's avatar
Anthony Larcher committed
296
        else:
Anthony Larcher's avatar
Anthony Larcher committed
297
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
298

Anthony Larcher's avatar
Anthony Larcher committed
299
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
300
301
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
302
303
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
304
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
305
306

    """
Anthony Larcher's avatar
Anthony Larcher committed
307
308
309
310
    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
311
    """
Anthony Larcher's avatar
Anthony Larcher committed
312
313
314
315
    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"])
316

Anthony Larcher's avatar
Anthony Larcher committed
317
    torch.manual_seed(dataset_params['seed'])
318
319
320
321
322
    training_set = SideSet(dataset_yaml, 
                           set_type="train", 
                           dataset_df=training_df, 
                           chunk_per_segment=dataset_params['chunk_per_segment'], 
                           overlap=dataset_params['overlap'])
Anthony Larcher's avatar
Anthony Larcher committed
323
324
325
    training_loader = DataLoader(training_set,
                                 batch_size=dataset_params["batch_size"],
                                 shuffle=True,
Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
326
                                 drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
327
                                 num_workers=num_thread)
328

Anthony Larcher's avatar
Anthony Larcher committed
329
330
331
    validation_set = SideSet(dataset_yaml, set_type="validation", dataset_df=validation_df)
    validation_loader = DataLoader(validation_set,
                                   batch_size=dataset_params["batch_size"],
Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
332
                                   drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
333
                                   num_workers=num_thread)
Anthony Larcher's avatar
Anthony Larcher committed
334
335
336
337
338

    """
    Set the training options
    """
    if type(model) is Xtractor:
Anthony Larcher's avatar
Anthony Larcher committed
339
        optimizer = torch.optim.SGD([
Anthony Larcher's avatar
Anthony Larcher committed
340
341
342
343
344
345
            {'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
346
            lr=lr, momentum=0.9
Anthony Larcher's avatar
Anthony Larcher committed
347
348
        )
    else:
Anthony Larcher's avatar
Anthony Larcher committed
349
        optimizer = torch.optim.SGD([
Anthony Larcher's avatar
Anthony Larcher committed
350
351
352
353
354
355
            {'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
356
            lr=lr, momentum=0.9
Anthony Larcher's avatar
Anthony Larcher committed
357
        )
Anthony Larcher's avatar
Anthony Larcher committed
358
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', verbose=True)
359

Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
360
    best_accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
361
    for epoch in range(1, epochs + 1):
362
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
363
        model = train_epoch(model, epoch, training_loader, optimizer, dataset_params["log_interval"], device=device, clipping=clipping)
364
365

        # Add the cross validation here
Anthony Larcher's avatar
Anthony Larcher committed
366
        accuracy, val_loss = cross_validation(model, validation_loader, device=device)
367
368
369
370
        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
371
        print(f"Learning rate is {optimizer.param_groups[0]['lr']}")
372

Anthony Larcher's avatar
Anthony Larcher committed
373
374
375
376
377
378
379
380
381
382
        # 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
Anthony Larcher's avatar
minor    
Anthony Larcher committed
383
        }, is_best, filename = tmp_model_name+".pt", best_filename=best_model_name+'.pt')
Anthony Larcher's avatar
Anthony Larcher committed
384
385
386

        if is_best:
            best_accuracy_epoch = epoch
387

Anthony Larcher's avatar
Anthony Larcher committed
388
    logging.critical(f"Best accuracy {best_accuracy * 100.} obtained at epoch {best_accuracy_epoch}")
389

Anthony Larcher's avatar
Anthony Larcher committed
390
def train_epoch(model, epoch, training_loader, optimizer, log_interval, device, clipping=False):
391
392
393
394
395
396
397
398
399
400
    """

    :param model:
    :param epoch:
    :param train_seg_df:
    :param speaker_dict:
    :param optimizer:
    :param args:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
401
    model.train()
Anthony Larcher's avatar
Anthony Larcher committed
402
    criterion = torch.nn.CrossEntropyLoss(reduction='mean')
403
404

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
405
    for batch_idx, (data, target) in enumerate(training_loader):
406
407
408
409
410
        target = target.squeeze()
        optimizer.zero_grad()
        output = model(data.to(device))
        loss = criterion(output, target.to(device))
        loss.backward()
Anthony Larcher's avatar
Anthony Larcher committed
411
412
        if clipping:
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
413
414
415
        optimizer.step()
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()

Anthony Larcher's avatar
Anthony Larcher committed
416
        if batch_idx % log_interval == 0:
Anthony Larcher's avatar
Anthony Larcher committed
417
            batch_size = target.shape[0]
418
            logging.critical('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
Anthony Larcher's avatar
Anthony Larcher committed
419
420
                epoch, batch_idx + 1, training_loader.__len__(),
                       100. * batch_idx / training_loader.__len__(), loss.item(),
Anthony Larcher's avatar
Anthony Larcher committed
421
                       100.0 * accuracy.item() / ((batch_idx + 1) * batch_size)))
422
423
424
    return model


Anthony Larcher's avatar
Anthony Larcher committed
425
def cross_validation(model, validation_loader, device):
426
427
428
    """

    :param model:
Anthony Larcher's avatar
Anthony Larcher committed
429
430
    :param validation_loader:
    :param device:
431
432
433
434
435
    :return:
    """
    model.eval()

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
436
    loss = 0.0
437
    criterion = torch.nn.CrossEntropyLoss()
Anthony Larcher's avatar
Anthony Larcher committed
438
439
440
441
442
443
    with torch.no_grad():
        for batch_idx, (data, target) in enumerate(validation_loader):
            batch_size = target.shape[0]
            target = target.squeeze()
            output = model(data.to(device))
            accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()
444

Anthony Larcher's avatar
Anthony Larcher committed
445
446
447
            loss += criterion(output, target.to(device))
    
    return 100. * accuracy.cpu().numpy() / ((batch_idx + 1) * batch_size), loss.cpu().numpy() / ((batch_idx + 1) * batch_size)
448
449


450
def extract_idmap(args, device_id, segment_indices, fs_params, idmap_name, output_queue):
Anthony Larcher's avatar
Anthony Larcher committed
451
    """
Anthony Larcher's avatar
Anthony Larcher committed
452
453
    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
454
    """
455
    # device = torch.device("cuda:{}".format(device_ID))
Anthony Larcher's avatar
Anthony Larcher committed
456
    device = torch.device('cpu')
Anthony Larcher's avatar
Anthony Larcher committed
457
458
459
460
461
462
463
464
465
466
467
468
469

    # 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
470
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
471
472
473
474
475
476
477
478
    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
479
480
481
482
483
484
    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
485
486
487
488
489
490

    # 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
491
        logging.critical('Process file {}, [{} / {}]'.format(segment_id, idx, segment_loader.__len__()))
Anthony Larcher's avatar
Anthony Larcher committed
492

Anthony Larcher's avatar
Anthony Larcher committed
493
494
495
        if list(data.shape)[2] < 20:
            pass
        else:
Anthony Larcher's avatar
Anthony Larcher committed
496
497
498
499
500
501
502
            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
503

Anthony Larcher's avatar
Anthony Larcher committed
504
    output_queue.put((segment_indices, emb_1, emb_2, emb_3, emb_4, emb_5, emb_6))
Anthony Larcher's avatar
Anthony Larcher committed
505
506


Anthony Larcher's avatar
Anthony Larcher committed
507
def extract_parallel(args, fs_params):
508
509
510
511
512
513
    """

    :param args:
    :param fs_params:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
514
515
516
    emb_a_size = 512
    emb_b_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
517
    idmap = IdMap(args.idmap)
Anthony Larcher's avatar
Anthony Larcher committed
518

Anthony Larcher's avatar
Anthony Larcher committed
519
520
521
522
523
524
525
526
527
528
529
530
531
    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
532
533
534

    # Split the indices
    mega_batch_size = idmap.leftids.shape[0] // args.num_processes
Anthony Larcher's avatar
Anthony Larcher committed
535
536
537

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

Anthony Larcher's avatar
Anthony Larcher committed
538
539
540
    segment_idx = []
    for ii in range(args.num_processes):
        segment_idx.append(
Anthony Larcher's avatar
Anthony Larcher committed
541
542
543
544
            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
545
546
547
548
549
550
551

    # 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
552
                       args=(args, rank, segment_idx[rank], fs_params, args.idmap, output_queue)
Anthony Larcher's avatar
Anthony Larcher committed
553
554
555
556
557
558
559
                       )
        # 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
560
561
562
563
564
565
566
        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
567
568
569
570

    for p in processes:
        p.join()

Anthony Larcher's avatar
Anthony Larcher committed
571
    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
572