xvector.py 26.5 KB
Newer Older
Anthony Larcher's avatar
Anthony Larcher committed
1
# coding: utf-8 -*-
Anthony Larcher's avatar
Anthony Larcher committed
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
#
# This file is part of SIDEKIT.
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is free software: you can redistribute it and/or modify
# it under the terms of the GNU LLesser General Public License as
# published by the Free Software Foundation, either version 3 of the License,
# or (at your option) any later version.
#
# SIDEKIT is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU Lesser General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with SIDEKIT.  If not, see <http://www.gnu.org/licenses/>.

"""
Anthony Larcher's avatar
Anthony Larcher committed
25
Copyright 2014-2020 Yevhenii Prokopalo, Anthony Larcher
Anthony Larcher's avatar
Anthony Larcher committed
26
"""
Anthony Larcher's avatar
Anthony Larcher committed
27

Anthony Larcher's avatar
Anthony Larcher committed
28
29
import logging
import numpy
Anthony Larcher's avatar
Anthony Larcher committed
30
import pandas
Anthony Larcher's avatar
minor    
Anthony Larcher committed
31
import pickle
Anthony Larcher's avatar
Anthony Larcher committed
32
import shutil
Anthony Larcher's avatar
Anthony Larcher committed
33
import torch
Anthony Larcher's avatar
Anthony Larcher committed
34
35
import torch.optim as optim
import torch.multiprocessing as mp
Anthony Larcher's avatar
Anthony Larcher committed
36
37
import yaml

Anthony Larcher's avatar
Anthony Larcher committed
38
from torchvision import transforms
Anthony Larcher's avatar
Anthony Larcher committed
39
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
40
from .xsets import XvectorMultiDataset, StatDataset, VoxDataset, SideSet
Anthony Larcher's avatar
Anthony Larcher committed
41
from .xsets import IdMapSet
Anthony Larcher's avatar
Anthony Larcher committed
42
from .xsets import FrequencyMask, CMVN, TemporalMask, MFCC
Anthony Larcher's avatar
Anthony Larcher committed
43
from .res_net import RawPreprocessor, ResBlockWFMS
Anthony Larcher's avatar
Anthony Larcher committed
44
45
from ..bosaris import IdMap
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
46
from torch.utils.data import DataLoader
Anthony Larcher's avatar
Anthony Larcher committed
47
from sklearn.model_selection import train_test_split
Anthony Larcher's avatar
Anthony Larcher committed
48
from .sincnet import SincNet, SincConv1d
Anthony Larcher's avatar
Anthony Larcher committed
49
from tqdm import tqdm
Anthony Larcher's avatar
Anthony Larcher committed
50

Anthony Larcher's avatar
Anthony Larcher committed
51
52
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
53
__copyright__ = "Copyright 2015-2020 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
54
55
56
57
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
58
59


60
def get_lr(optimizer):
Anthony Larcher's avatar
Anthony Larcher committed
61
62
63
64
65
    """

    :param optimizer:
    :return:
    """
66
67
68
69
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
70
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
71
72
73
74
75
76
77
78
    """

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

Anthony Larcher's avatar
Anthony Larcher committed
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
class MeanStdPooling(torch.nn.Module):
    """
    Mean and Standard deviation pooling
    """
    def __init__(self):
        """

        """
        super(MeanStdPooling, self).__init__()
        pass

    def forward(self, x):
        """

        :param x:
        :return:
        """
        mean = torch.mean(x, dim=2)
        std = torch.std(x, dim=2)
        return torch.cat([mean, std], dim=1)
Anthony Larcher's avatar
Anthony Larcher committed
103

Anthony Larcher's avatar
Anthony Larcher committed
104
105
106
107
108
109
110
111
112
113
114
class GruPooling(torch.nn.Module):
    """

    """
    def __init__(self, input_size, gru_node, nb_gru_layer):
        """

        :param input_size:
        :param gru_node:
        :param nb_gru_layer:
        """
Anthony Larcher's avatar
Anthony Larcher committed
115
        super(GruPooling, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
        self.lrelu_keras = torch.nn.LeakyReLU(negative_slope = 0.3)
        self.bn_before_gru = torch.nn.BatchNorm1d(num_features = input_size)
        self.gru = torch.nn.GRU(input_size = input_size,
                          hidden_size = gru_node,
                          num_layers = nb_gru_layer,
                          batch_first = True)

    def forward(self, x):
        """

        :param x:
        :return:
        """
        x = self.bn_before_gru(x)
        x = self.lrelu_keras(x)
        x = x.permute(0, 2, 1)  #(batch, filt, time) >> (batch, time, filt)
        self.gru.flatten_parameters()
        x, _ = self.gru(x)
        x = x[:,-1,:]

        return x

Anthony Larcher's avatar
Anthony Larcher committed
138
class Xtractor(torch.nn.Module):
139
140
141
    """
    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
142

Anthony Larcher's avatar
Anthony Larcher committed
143
    def __init__(self, speaker_number, model_archi="xvector", norm_embedding=False):
Anthony Larcher's avatar
Anthony Larcher committed
144
145
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
146
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
147
        """
Anthony Larcher's avatar
Anthony Larcher committed
148
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
149
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
150
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
151
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
152

Anthony Larcher's avatar
Anthony Larcher committed
153
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
154
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
155
156
157
            self.activation = torch.nn.LeakyReLU(0.2)

            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
158

Anthony Larcher's avatar
xv    
Anthony Larcher committed
159
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
160
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
161
162
163
164
165
166
167
168
                ("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
169
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
170
171
                ("activation4", torch.nn.LeakyReLU(0.2)),
                ("norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
172
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
173
174
175
176
                ("activation5", torch.nn.LeakyReLU(0.2)),
                ("norm5", torch.nn.BatchNorm1d(1536))
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
177
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
178
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
179
180
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
181
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
182
183
                ("activation6", torch.nn.LeakyReLU(0.2)),
                ("norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
184
                ("dropout6", torch.nn.Dropout(p=0.05)),
Anthony Larcher's avatar
Anthony Larcher committed
185
                ("linear7", torch.nn.Linear(512, 512)),
Anthony Larcher's avatar
Anthony Larcher committed
186
187
                ("activation7", torch.nn.LeakyReLU(0.2)),
                ("norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
188
                ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
Anthony Larcher's avatar
Anthony Larcher committed
189
190
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
191
192
193
194
            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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
        elif model_archi == "rawnet2":
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

            self.preprocessor = RawPreprocessor(nb_samp=48000,
                                                in_channels=1,
                                                filts=filts[0],
                                                first_conv=3)

            self.sequence_network = torch.nn.Sequential(OrderedDict([
                ("block0", ResBlockWFMS(nb_filts=filts[1], first=True)),
                ("block1", ResBlockWFMS(nb_filts=filts[1])),
                ("block2", ResBlockWFMS(nb_filts=filts[2])),
                ("block3", ResBlockWFMS(nb_filts=[filts[2][1], filts[2][1]])),
                ("block4", ResBlockWFMS(nb_filts=[filts[2][1], filts[2][1]])),
                ("block5", ResBlockWFMS(nb_filts=[filts[2][1], filts[2][1]]))
            ]))

            self.stat_pooling = GruPooling(input_size=filts[2][-1],
                                           gru_node=1024,
                                           nb_gru_layer=1)

            self.before_speaker_embedding = torch.nn.Linear(in_features = 1024,
                                                            out_features = 1024)

            self.after_speaker_embedding = torch.nn.Linear(in_features = 1024,
                                                           out_features = int(self.speaker_number),
                                                           bias = True)

Anthony Larcher's avatar
Anthony Larcher committed
224
225
        else:
            # Load Yaml configuration
Anthony Larcher's avatar
Anthony Larcher committed
226
            with open(model_archi, 'r') as fh:
Anthony Larcher's avatar
Anthony Larcher committed
227
228
                cfg = yaml.load(fh, Loader=yaml.FullLoader)

Anthony Larcher's avatar
Anthony Larcher committed
229
230
231
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
232
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
233
234
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
235
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
236
237
238
239
240
241
242
243
244
245
246
247
                        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
248
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
249
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
250
251
252
253
                    self.preprocessor = RawPreprocessor(nb_samp=48000,
                                                        in_channels=1,
                                                        filts=128,
                                                        first_conv=3)
Anthony Larcher's avatar
Anthony Larcher committed
254
255

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
256
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
257
            """
Anthony Larcher's avatar
Anthony Larcher committed
258
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
259
260
261
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
            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
278
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
279
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
280
281
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
282
283
284
285
286
287
288
289
                    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
290
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
291
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
292

Anthony Larcher's avatar
Anthony Larcher committed
293
294
295
            """
            Prepapre last part of the network (after pooling)
            """
Anthony Larcher's avatar
Anthony Larcher committed
296
297
            # Create sequential object for the second part of the network
            input_size = input_size * 2
Anthony Larcher's avatar
xv    
Anthony Larcher committed
298
299
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
300
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
301
302
                    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
303
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
304
305
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
306
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
307
308

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
309
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
310
311

                elif k.startswith('norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
312
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
313
314

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
315
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
316

Anthony Larcher's avatar
Anthony Larcher committed
317
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
318
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
319
320
321
322
323

            # 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
324
325
                    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
326
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
327
328
                        after_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
329
                        input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
330
331
332
333
334
335
336
337

                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
338
                    after_embedding_layers.append((k, torch.nn.Dropout(p=cfg["after_embedding"][k])))
Anthony Larcher's avatar
Anthony Larcher committed
339

Anthony Larcher's avatar
Anthony Larcher committed
340
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict(after_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
341
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
342

Anthony Larcher's avatar
Anthony Larcher committed
343
    def forward(self, x, is_eval=False):
344
345
346
        """

        :param x:
Anthony Larcher's avatar
Anthony Larcher committed
347
        :param is_eval:
348
349
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
350
351
352
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
353
        x = self.sequence_network(x)
354

Anthony Larcher's avatar
Anthony Larcher committed
355
356
357
358
359
360
361
362
        # 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
363

Anthony Larcher's avatar
Anthony Larcher committed
364
365
366
367
        if self.norm_embedding:
            x_norm = x.norm(p=2,dim=1, keepdim=True) / 10.
            x = torch.div(x, x_norm)

Anthony Larcher's avatar
Anthony Larcher committed
368
369
        x = self.after_speaker_embedding(x)
        return x
Anthony Larcher's avatar
Anthony Larcher committed
370

Anthony Larcher's avatar
Anthony Larcher committed
371

Anthony Larcher's avatar
Anthony Larcher committed
372
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
373
           dataset_yaml,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
374
           epochs=100,
Anthony Larcher's avatar
Anthony Larcher committed
375
           lr=0.01,
Anthony Larcher's avatar
Anthony Larcher committed
376
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
377
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
378
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
379
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
380
           multi_gpu=True,
381
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
382
           opt='sgd',
Anthony Larcher's avatar
Anthony Larcher committed
383
           num_thread=1):
384
385
    """

Anthony Larcher's avatar
Anthony Larcher committed
386
387
388
389
390
391
392
393
394
395
396
    :param speaker_number:
    :param dataset_yaml:
    :param epochs:
    :param lr:
    :param model_yaml:
    :param model_name:
    :param tmp_model_name:
    :param best_model_name:
    :param multi_gpu:
    :param clipping:
    :param num_thread:
397
398
    :return:
    """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
399
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
400

401
    # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
402
403
404
405
406
407
408
    if model_name is not None:
        # Load the model
        logging.critical(f"*** Load model from = {model_name}")
        checkpoint = torch.load(model_name)
        model = Xtractor(speaker_number, model_yaml)
        model.load_state_dict(checkpoint["model_state_dict"])
    else:
Anthony Larcher's avatar
Anthony Larcher committed
409
410
        # Initialize a first model
        if model_yaml is None:
Anthony Larcher's avatar
Anthony Larcher committed
411
            model = Xtractor(speaker_number)
Anthony Larcher's avatar
Anthony Larcher committed
412
        else:
Anthony Larcher's avatar
Anthony Larcher committed
413
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
414

Anthony Larcher's avatar
Anthony Larcher committed
415
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
416
417
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
418
419
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
420
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
421
422

    """
Anthony Larcher's avatar
Anthony Larcher committed
423
424
425
426
    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
427
    """
Anthony Larcher's avatar
Anthony Larcher committed
428
429
430
431
    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"])
432

Anthony Larcher's avatar
Anthony Larcher committed
433
    torch.manual_seed(dataset_params['seed'])
434
435
436
    training_set = SideSet(dataset_yaml, 
                           set_type="train", 
                           dataset_df=training_df, 
Anthony Larcher's avatar
Anthony Larcher committed
437
438
                           chunk_per_segment=dataset_params['train']['chunk_per_segment'], 
                           overlap=dataset_params['train']['overlap'])
Anthony Larcher's avatar
Anthony Larcher committed
439
440
441
    training_loader = DataLoader(training_set,
                                 batch_size=dataset_params["batch_size"],
                                 shuffle=True,
Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
442
                                 drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
443
                                 num_workers=num_thread)
444

Anthony Larcher's avatar
Anthony Larcher committed
445
446
447
    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
448
                                   drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
449
                                   num_workers=num_thread)
Anthony Larcher's avatar
Anthony Larcher committed
450
451
452
453

    """
    Set the training options
    """
Anthony Larcher's avatar
Anthony Larcher committed
454
455
    if opt == 'sgd':
        _optimizer = torch.optim.SGD
Anthony Larcher's avatar
Anthony Larcher committed
456
        _options = {'lr': lr, 'momentum': 0.9}
Anthony Larcher's avatar
Anthony Larcher committed
457
458
    elif opt == 'adam':
        _optimizer = torch.optim.Adam
Anthony Larcher's avatar
Anthony Larcher committed
459
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
460
461
    elif opt == 'rmsprop':
        _optimizer = torch.optim.RMSprop
Anthony Larcher's avatar
Anthony Larcher committed
462
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
463

Anthony Larcher's avatar
Anthony Larcher committed
464
    if type(model) is Xtractor:
Anthony Larcher's avatar
Anthony Larcher committed
465
        optimizer = _optimizer([
Anthony Larcher's avatar
Anthony Larcher committed
466
467
468
469
470
471
            {'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
472
            **_options
Anthony Larcher's avatar
Anthony Larcher committed
473
474
        )
    else:
Anthony Larcher's avatar
Anthony Larcher committed
475
        optimizer = _optimizer([
Anthony Larcher's avatar
Anthony Larcher committed
476
477
478
479
480
481
            {'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
482
            **_options
Anthony Larcher's avatar
Anthony Larcher committed
483
        )
Anthony Larcher's avatar
Anthony Larcher committed
484
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', verbose=True)
485

Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
486
    best_accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
487
    best_accuracy_epoch = 1
Anthony Larcher's avatar
Anthony Larcher committed
488
    for epoch in range(1, epochs + 1):
489
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
490
491
492
493
494
495
496
        model = train_epoch(model,
                            epoch,
                            training_loader,
                            optimizer,
                            dataset_params["log_interval"],
                            device=device,
                            clipping=clipping)
497
498

        # Add the cross validation here
Anthony Larcher's avatar
Anthony Larcher committed
499
        accuracy, val_loss = cross_validation(model, validation_loader, device=device)
500
501
502
503
        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
504
        print(f"Learning rate is {optimizer.param_groups[0]['lr']}")
505

Anthony Larcher's avatar
Anthony Larcher committed
506
507
508
509
        # remember best accuracy and save checkpoint
        is_best = accuracy > best_accuracy
        best_accuracy = max(accuracy, best_accuracy)

Anthony Larcher's avatar
Anthony Larcher committed
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
        if type(model) is Xtractor:
            save_checkpoint({
                'epoch': epoch,
                'model_state_dict': model.state_dict(),
                'optimizer_state_dict': optimizer.state_dict(),
                'accuracy': best_accuracy,
                'scheduler': scheduler
            }, is_best, filename=tmp_model_name+".pt", best_filename=best_model_name+'.pt')
        else:
            save_checkpoint({
                'epoch': epoch,
                'model_state_dict': model.module.state_dict(),
                'optimizer_state_dict': optimizer.state_dict(),
                'accuracy': best_accuracy,
                'scheduler': scheduler
            }, is_best, filename=tmp_model_name+".pt", best_filename=best_model_name+'.pt')
Anthony Larcher's avatar
Anthony Larcher committed
526
527
528

        if is_best:
            best_accuracy_epoch = epoch
529

Anthony Larcher's avatar
Anthony Larcher committed
530
    logging.critical(f"Best accuracy {best_accuracy * 100.} obtained at epoch {best_accuracy_epoch}")
531

Anthony Larcher's avatar
Anthony Larcher committed
532

Anthony Larcher's avatar
Anthony Larcher committed
533
def train_epoch(model, epoch, training_loader, optimizer, log_interval, device, clipping=False):
534
535
536
537
    """

    :param model:
    :param epoch:
Anthony Larcher's avatar
Anthony Larcher committed
538
    :param training_loader:
539
    :param optimizer:
Anthony Larcher's avatar
Anthony Larcher committed
540
541
542
    :param log_interval:
    :param device:
    :param clipping:
543
544
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
545
    model.train()
Anthony Larcher's avatar
Anthony Larcher committed
546
    criterion = torch.nn.CrossEntropyLoss(reduction='mean')
547
548

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
549
    for batch_idx, (data, target) in enumerate(training_loader):
550
551
552
553
554
        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
555
556
        if clipping:
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
557
558
559
        optimizer.step()
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()

Anthony Larcher's avatar
Anthony Larcher committed
560
        if batch_idx % log_interval == 0:
Anthony Larcher's avatar
Anthony Larcher committed
561
            batch_size = target.shape[0]
562
            logging.critical('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
Anthony Larcher's avatar
Anthony Larcher committed
563
                epoch, batch_idx + 1, training_loader.__len__(),
Anthony Larcher's avatar
Anthony Larcher committed
564
565
                100. * batch_idx / training_loader.__len__(), loss.item(),
                100.0 * accuracy.item() / ((batch_idx + 1) * batch_size)))
566
567
568
    return model


Anthony Larcher's avatar
Anthony Larcher committed
569
def cross_validation(model, validation_loader, device):
570
571
572
    """

    :param model:
Anthony Larcher's avatar
Anthony Larcher committed
573
574
    :param validation_loader:
    :param device:
575
576
577
578
579
    :return:
    """
    model.eval()

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
580
    loss = 0.0
581
    criterion = torch.nn.CrossEntropyLoss()
Anthony Larcher's avatar
Anthony Larcher committed
582
583
584
585
586
587
    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()
588

Anthony Larcher's avatar
Anthony Larcher committed
589
590
            loss += criterion(output, target.to(device))
    
Anthony Larcher's avatar
Anthony Larcher committed
591
592
593
594
    return 100. * accuracy.cpu().numpy() / ((batch_idx + 1) * batch_size), \
           loss.cpu().numpy() / ((batch_idx + 1) * batch_size)


Anthony Larcher's avatar
Anthony Larcher committed
595
596
597
598
599
600
601
602
def extract_embeddings(idmap_name,
                       speaker_number,
                       model_filename,
                       model_yaml,
                       data_root_name ,
                       device,
                       file_extension="wav",
                       transform_pipeline=None):
Anthony Larcher's avatar
Anthony Larcher committed
603

Anthony Larcher's avatar
Anthony Larcher committed
604
    if isinstance(idmap_name, IdMap):
605
606
607
608
        idmap = idmap_name
    else:
        idmap = IdMap(idmap_name)

Anthony Larcher's avatar
Anthony Larcher committed
609
    # Create dataset to load the data
Anthony Larcher's avatar
Anthony Larcher committed
610
611
612
613
    dataset = IdMapSet(idmap_name=idmap_name,
                       data_root_path=data_root_name,
                       file_extension=file_extension,
                       transform_pipeline=transform_pipeline)
Anthony Larcher's avatar
Anthony Larcher committed
614
615

    # Load the model
616
617
618
619
620
621
622
    if isinstance(model_filename, str):
        checkpoint = torch.load(model_filename)
        model = Xtractor(speaker_number, model_archi=model_yaml)
        model.load_state_dict(checkpoint["model_state_dict"])
    else:
        model = model_filename

Anthony Larcher's avatar
Anthony Larcher committed
623
624
    model.eval()
    model.to(device)
625

Anthony Larcher's avatar
Anthony Larcher committed
626
627
628
    # Get the size of embeddings to extract
    name = list(model.before_speaker_embedding.state_dict().keys())[-1].split('.')[0] + '.weight'
    emb_size = model.before_speaker_embedding.state_dict()[name].shape[0]
Anthony Larcher's avatar
Anthony Larcher committed
629
    
Anthony Larcher's avatar
Anthony Larcher committed
630
    # Create the StatServer
Anthony Larcher's avatar
Anthony Larcher committed
631
    embeddings = StatServer()
Anthony Larcher's avatar
Anthony Larcher committed
632
633
634
635
636
637
    embeddings.modelset = idmap.leftids
    embeddings.segset = idmap.rightids
    embeddings.start = idmap.start
    embeddings.stop = idmap.stop
    embeddings.stat0 = numpy.ones((embeddings.modelset.shape[0], 1))
    embeddings.stat1 = numpy.ones((embeddings.modelset.shape[0], emb_size))
Anthony Larcher's avatar
Anthony Larcher committed
638

Anthony Larcher's avatar
Anthony Larcher committed
639
640
    # Process the data
    with torch.no_grad():
Anthony Larcher's avatar
Anthony Larcher committed
641
        for idx in tqdm(range(len(dataset))):
Anthony Larcher's avatar
Anthony Larcher committed
642
            data, mod, seg, start, stop = dataset[idx]
Anthony Larcher's avatar
Anthony Larcher committed
643
            vec = model(data[None, :, :].to(device), is_eval=True)
Anthony Larcher's avatar
Anthony Larcher committed
644
645
646
647
648
649
            #current_idx = numpy.argwhere(numpy.logical_and(idmap.leftids == mod, idmap.rightids == seg))[0][0]
            embeddings.start[idx] = start
            embeddings.stop[idx] = stop
            embeddings.modelset[idx] = mod
            embeddings.segset[idx] = seg
            embeddings.stat1[idx, :] = vec.detach().cpu()
Anthony Larcher's avatar
Anthony Larcher committed
650
651
652
653

    return embeddings


Anthony Larcher's avatar
Anthony Larcher committed
654
655
656
657
658
659
660
661
662
663
664
665
def extract_sliding_embedding(idmap_name,
                              window_length,
                              sample_rate,
                              overlap,
                              speaker_number,
                              model_filename,
                              model_yaml,
                              data_root_name ,
                              device,
                              file_extension="wav",
                              transform_pipeline=None):

666
667
668
669
670
671
672
673
674
675
676
677
678
679

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

    # Create temporary lists
    nb_chunks = 0
    model_names = []
    segment_names = []
    starts = []
    stops = []
    for mod, seg, start, stop in zip(input_idmap.leftids, input_idmap.rightids, input_idmap.start, input_idmap.stop):
        # Compute the number of chunks to process
        chunk_starts = numpy.arange(start,
                                    stop - int(sample_rate * window_length),
Anthony Larcher's avatar
Anthony Larcher committed
680
                                    int(sample_rate * (window_length - overlap)))
681
682

        # Create a numpy array to store the current x-vectors
Anthony Larcher's avatar
Anthony Larcher committed
683
        model_names.append(numpy.array([mod + f"_{ii}" for ii in range(len(chunk_starts))]).astype("U"))
684
685
686
687
688
689
690
        segment_names.append(numpy.array([seg, ] * chunk_starts.shape[0]))
        starts.append(chunk_starts)
        stops.append(chunk_starts + sample_rate * window_length)

        nb_chunks += len(chunk_starts)

    sliding_idmap = IdMap()
Anthony Larcher's avatar
Anthony Larcher committed
691
692
693
694
    sliding_idmap.leftids = numpy.hstack(model_names)
    sliding_idmap.rightids = numpy.hstack(segment_names)
    sliding_idmap.start = numpy.hstack(starts)
    sliding_idmap.stop = numpy.hstack(stops)
695
    assert sliding_idmap.validate()
Anthony Larcher's avatar
Anthony Larcher committed
696

Anthony Larcher's avatar
Anthony Larcher committed
697
698
699
700
701
702
    embeddings = extract_embeddings(sliding_idmap,
                                 speaker_number,
                                 model_filename,
                                 model_yaml,
                                 data_root_name,
                                 device,
Anthony Larcher's avatar
Anthony Larcher committed
703
704
                                 file_extension=file_extension,
                                 transform_pipeline=transform_pipeline)
Anthony Larcher's avatar
Anthony Larcher committed
705
706

    return embeddings
Anthony Larcher's avatar
Anthony Larcher committed
707
708