xvector.py 25.8 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
83


Anthony Larcher's avatar
Anthony Larcher committed
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
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:
        """
        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
117
class Xtractor(torch.nn.Module):
118
119
120
    """
    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
121

Anthony Larcher's avatar
Anthony Larcher committed
122
    def __init__(self, speaker_number, model_archi="xvector", norm_embedding=False):
Anthony Larcher's avatar
Anthony Larcher committed
123
124
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
125
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
126
        """
Anthony Larcher's avatar
Anthony Larcher committed
127
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
128
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
129
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
130
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
131

Anthony Larcher's avatar
Anthony Larcher committed
132
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
133
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
134
135
136
            self.activation = torch.nn.LeakyReLU(0.2)

            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
137

Anthony Larcher's avatar
xv    
Anthony Larcher committed
138
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
139
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
140
141
142
143
144
145
146
147
                ("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
148
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
149
150
                ("activation4", torch.nn.LeakyReLU(0.2)),
                ("norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
151
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
152
153
154
155
                ("activation5", torch.nn.LeakyReLU(0.2)),
                ("norm5", torch.nn.BatchNorm1d(1536))
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
156
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
157
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
158
159
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
160
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
161
162
                ("activation6", torch.nn.LeakyReLU(0.2)),
                ("norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
163
                ("dropout6", torch.nn.Dropout(p=0.05)),
Anthony Larcher's avatar
Anthony Larcher committed
164
                ("linear7", torch.nn.Linear(512, 512)),
Anthony Larcher's avatar
Anthony Larcher committed
165
166
                ("activation7", torch.nn.LeakyReLU(0.2)),
                ("norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
167
                ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
Anthony Larcher's avatar
Anthony Larcher committed
168
169
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
170
171
172
173
            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
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
        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
203
204
        else:
            # Load Yaml configuration
Anthony Larcher's avatar
Anthony Larcher committed
205
            with open(model_archi, 'r') as fh:
Anthony Larcher's avatar
Anthony Larcher committed
206
207
                cfg = yaml.load(fh, Loader=yaml.FullLoader)

Anthony Larcher's avatar
Anthony Larcher committed
208
209
210
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
211
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
212
213
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
214
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
215
216
217
218
219
220
221
222
223
224
225
226
                        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
227
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
228
229
                elif cfg['preprocessor']["type"] == "rawnet2":
                    self.preprocessor =
Anthony Larcher's avatar
Anthony Larcher committed
230
231

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
232
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
233
            """
Anthony Larcher's avatar
Anthony Larcher committed
234
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
235
236
237
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
            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
254
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
255
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
256
257
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
258
259
260
261
262
263
264
265
                    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
266
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
267
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
268

Anthony Larcher's avatar
Anthony Larcher committed
269
270
271
            """
            Prepapre last part of the network (after pooling)
            """
Anthony Larcher's avatar
Anthony Larcher committed
272
273
            # Create sequential object for the second part of the network
            input_size = input_size * 2
Anthony Larcher's avatar
xv    
Anthony Larcher committed
274
275
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
276
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
277
278
                    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
279
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
280
281
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
282
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
283
284

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
285
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
286
287

                elif k.startswith('norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
288
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
289
290

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
291
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
292

Anthony Larcher's avatar
Anthony Larcher committed
293
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
294
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
295
296
297
298
299

            # 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
300
301
                    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
302
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
303
304
                        after_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
305
                        input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
306
307
308
309
310
311
312
313

                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
314
                    after_embedding_layers.append((k, torch.nn.Dropout(p=cfg["after_embedding"][k])))
Anthony Larcher's avatar
Anthony Larcher committed
315

Anthony Larcher's avatar
Anthony Larcher committed
316
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict(after_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
317
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
318

Anthony Larcher's avatar
Anthony Larcher committed
319
    def forward(self, x, is_eval=False):
320
321
322
        """

        :param x:
Anthony Larcher's avatar
Anthony Larcher committed
323
        :param is_eval:
324
325
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
326
327
328
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
329
        x = self.sequence_network(x)
330

Anthony Larcher's avatar
Anthony Larcher committed
331
332
333
334
335
336
337
338
        # 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
339

Anthony Larcher's avatar
Anthony Larcher committed
340
341
342
343
        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
344
345
        x = self.after_speaker_embedding(x)
        return x
Anthony Larcher's avatar
Anthony Larcher committed
346

Anthony Larcher's avatar
Anthony Larcher committed
347

Anthony Larcher's avatar
Anthony Larcher committed
348
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
349
           dataset_yaml,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
350
           epochs=100,
Anthony Larcher's avatar
Anthony Larcher committed
351
           lr=0.01,
Anthony Larcher's avatar
Anthony Larcher committed
352
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
353
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
354
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
355
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
356
           multi_gpu=True,
357
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
358
           opt='sgd',
Anthony Larcher's avatar
Anthony Larcher committed
359
           num_thread=1):
360
361
    """

Anthony Larcher's avatar
Anthony Larcher committed
362
363
364
365
366
367
368
369
370
371
372
    :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:
373
374
    :return:
    """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
375
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
376

377
    # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
378
379
380
381
382
383
384
    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
385
386
        # Initialize a first model
        if model_yaml is None:
Anthony Larcher's avatar
Anthony Larcher committed
387
            model = Xtractor(speaker_number)
Anthony Larcher's avatar
Anthony Larcher committed
388
        else:
Anthony Larcher's avatar
Anthony Larcher committed
389
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
390

Anthony Larcher's avatar
Anthony Larcher committed
391
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
392
393
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
394
395
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
396
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
397
398

    """
Anthony Larcher's avatar
Anthony Larcher committed
399
400
401
402
    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
403
    """
Anthony Larcher's avatar
Anthony Larcher committed
404
405
406
407
    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"])
408

Anthony Larcher's avatar
Anthony Larcher committed
409
    torch.manual_seed(dataset_params['seed'])
410
411
412
    training_set = SideSet(dataset_yaml, 
                           set_type="train", 
                           dataset_df=training_df, 
Anthony Larcher's avatar
Anthony Larcher committed
413
414
                           chunk_per_segment=dataset_params['train']['chunk_per_segment'], 
                           overlap=dataset_params['train']['overlap'])
Anthony Larcher's avatar
Anthony Larcher committed
415
416
417
    training_loader = DataLoader(training_set,
                                 batch_size=dataset_params["batch_size"],
                                 shuffle=True,
Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
418
                                 drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
419
                                 num_workers=num_thread)
420

Anthony Larcher's avatar
Anthony Larcher committed
421
422
423
    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
424
                                   drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
425
                                   num_workers=num_thread)
Anthony Larcher's avatar
Anthony Larcher committed
426
427
428
429

    """
    Set the training options
    """
Anthony Larcher's avatar
Anthony Larcher committed
430
431
    if opt == 'sgd':
        _optimizer = torch.optim.SGD
Anthony Larcher's avatar
Anthony Larcher committed
432
        _options = {'lr': lr, 'momentum': 0.9}
Anthony Larcher's avatar
Anthony Larcher committed
433
434
    elif opt == 'adam':
        _optimizer = torch.optim.Adam
Anthony Larcher's avatar
Anthony Larcher committed
435
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
436
437
    elif opt == 'rmsprop':
        _optimizer = torch.optim.RMSprop
Anthony Larcher's avatar
Anthony Larcher committed
438
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
439

Anthony Larcher's avatar
Anthony Larcher committed
440
    if type(model) is Xtractor:
Anthony Larcher's avatar
Anthony Larcher committed
441
        optimizer = _optimizer([
Anthony Larcher's avatar
Anthony Larcher committed
442
443
444
445
446
447
            {'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
448
            **_options
Anthony Larcher's avatar
Anthony Larcher committed
449
450
        )
    else:
Anthony Larcher's avatar
Anthony Larcher committed
451
        optimizer = _optimizer([
Anthony Larcher's avatar
Anthony Larcher committed
452
453
454
455
456
457
            {'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
458
            **_options
Anthony Larcher's avatar
Anthony Larcher committed
459
        )
Anthony Larcher's avatar
Anthony Larcher committed
460
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', verbose=True)
461

Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
462
    best_accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
463
    best_accuracy_epoch = 1
Anthony Larcher's avatar
Anthony Larcher committed
464
    for epoch in range(1, epochs + 1):
465
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
466
467
468
469
470
471
472
        model = train_epoch(model,
                            epoch,
                            training_loader,
                            optimizer,
                            dataset_params["log_interval"],
                            device=device,
                            clipping=clipping)
473
474

        # Add the cross validation here
Anthony Larcher's avatar
Anthony Larcher committed
475
        accuracy, val_loss = cross_validation(model, validation_loader, device=device)
476
477
478
479
        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
480
        print(f"Learning rate is {optimizer.param_groups[0]['lr']}")
481

Anthony Larcher's avatar
Anthony Larcher committed
482
483
484
485
        # remember best accuracy and save checkpoint
        is_best = accuracy > best_accuracy
        best_accuracy = max(accuracy, best_accuracy)

Anthony Larcher's avatar
Anthony Larcher committed
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
        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
502
503
504

        if is_best:
            best_accuracy_epoch = epoch
505

Anthony Larcher's avatar
Anthony Larcher committed
506
    logging.critical(f"Best accuracy {best_accuracy * 100.} obtained at epoch {best_accuracy_epoch}")
507

Anthony Larcher's avatar
Anthony Larcher committed
508

Anthony Larcher's avatar
Anthony Larcher committed
509
def train_epoch(model, epoch, training_loader, optimizer, log_interval, device, clipping=False):
510
511
512
513
    """

    :param model:
    :param epoch:
Anthony Larcher's avatar
Anthony Larcher committed
514
    :param training_loader:
515
    :param optimizer:
Anthony Larcher's avatar
Anthony Larcher committed
516
517
518
    :param log_interval:
    :param device:
    :param clipping:
519
520
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
521
    model.train()
Anthony Larcher's avatar
Anthony Larcher committed
522
    criterion = torch.nn.CrossEntropyLoss(reduction='mean')
523
524

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
525
    for batch_idx, (data, target) in enumerate(training_loader):
526
527
528
529
530
        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
531
532
        if clipping:
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
533
534
535
        optimizer.step()
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()

Anthony Larcher's avatar
Anthony Larcher committed
536
        if batch_idx % log_interval == 0:
Anthony Larcher's avatar
Anthony Larcher committed
537
            batch_size = target.shape[0]
538
            logging.critical('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
Anthony Larcher's avatar
Anthony Larcher committed
539
                epoch, batch_idx + 1, training_loader.__len__(),
Anthony Larcher's avatar
Anthony Larcher committed
540
541
                100. * batch_idx / training_loader.__len__(), loss.item(),
                100.0 * accuracy.item() / ((batch_idx + 1) * batch_size)))
542
543
544
    return model


Anthony Larcher's avatar
Anthony Larcher committed
545
def cross_validation(model, validation_loader, device):
546
547
548
    """

    :param model:
Anthony Larcher's avatar
Anthony Larcher committed
549
550
    :param validation_loader:
    :param device:
551
552
553
554
555
    :return:
    """
    model.eval()

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
556
    loss = 0.0
557
    criterion = torch.nn.CrossEntropyLoss()
Anthony Larcher's avatar
Anthony Larcher committed
558
559
560
561
562
563
    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()
564

Anthony Larcher's avatar
Anthony Larcher committed
565
566
            loss += criterion(output, target.to(device))
    
Anthony Larcher's avatar
Anthony Larcher committed
567
568
569
570
    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
571
572
573
574
575
576
577
578
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
579

Anthony Larcher's avatar
Anthony Larcher committed
580
    if isinstance(idmap_name, IdMap):
581
582
583
584
        idmap = idmap_name
    else:
        idmap = IdMap(idmap_name)

Anthony Larcher's avatar
Anthony Larcher committed
585
    # Create dataset to load the data
Anthony Larcher's avatar
Anthony Larcher committed
586
587
588
589
    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
590
591

    # Load the model
592
593
594
595
596
597
598
    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
599
600
    model.eval()
    model.to(device)
601

Anthony Larcher's avatar
Anthony Larcher committed
602
603
604
    # 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
605
    
Anthony Larcher's avatar
Anthony Larcher committed
606
    # Create the StatServer
Anthony Larcher's avatar
Anthony Larcher committed
607
    embeddings = StatServer()
Anthony Larcher's avatar
Anthony Larcher committed
608
609
610
611
612
613
    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
614

Anthony Larcher's avatar
Anthony Larcher committed
615
616
    # Process the data
    with torch.no_grad():
Anthony Larcher's avatar
Anthony Larcher committed
617
        for idx in tqdm(range(len(dataset))):
Anthony Larcher's avatar
Anthony Larcher committed
618
            data, mod, seg, start, stop = dataset[idx]
Anthony Larcher's avatar
Anthony Larcher committed
619
            vec = model(data[None, :, :].to(device), is_eval=True)
Anthony Larcher's avatar
Anthony Larcher committed
620
621
622
623
624
625
            #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
626
627
628
629

    return embeddings


Anthony Larcher's avatar
Anthony Larcher committed
630
631
632
633
634
635
636
637
638
639
640
641
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):

642
643
644
645
646
647
648
649
650
651
652
653
654
655

    # 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
656
                                    int(sample_rate * (window_length - overlap)))
657
658

        # Create a numpy array to store the current x-vectors
Anthony Larcher's avatar
Anthony Larcher committed
659
        model_names.append(numpy.array([mod + f"_{ii}" for ii in range(len(chunk_starts))]).astype("U"))
660
661
662
663
664
665
666
        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
667
668
669
670
    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)
671
    assert sliding_idmap.validate()
Anthony Larcher's avatar
Anthony Larcher committed
672

Anthony Larcher's avatar
Anthony Larcher committed
673
674
675
676
677
678
    embeddings = extract_embeddings(sliding_idmap,
                                 speaker_number,
                                 model_filename,
                                 model_yaml,
                                 data_root_name,
                                 device,
Anthony Larcher's avatar
Anthony Larcher committed
679
680
                                 file_extension=file_extension,
                                 transform_pipeline=transform_pipeline)
Anthony Larcher's avatar
Anthony Larcher committed
681
682

    return embeddings
Anthony Larcher's avatar
Anthony Larcher committed
683
684