xvector.py 29.4 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
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


Anthony Larcher's avatar
Anthony Larcher committed
60
61
logging.basicConfig(format='%(asctime)s %(message)s')

62
def get_lr(optimizer):
Anthony Larcher's avatar
Anthony Larcher committed
63
64
65
66
67
    """

    :param optimizer:
    :return:
    """
68
69
70
71
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
72
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
73
74
75
76
77
78
79
80
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
81
82
83
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
84

Anthony Larcher's avatar
Anthony Larcher committed
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
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
105

Anthony Larcher's avatar
Anthony Larcher committed
106

Anthony Larcher's avatar
Anthony Larcher committed
107
108
109
110
111
112
113
114
115
116
117
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
118
        super(GruPooling, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
        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
141

Anthony Larcher's avatar
Anthony Larcher committed
142
class Xtractor(torch.nn.Module):
143
144
145
    """
    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
146

Anthony Larcher's avatar
Anthony Larcher committed
147
    def __init__(self, speaker_number, model_archi="xvector", norm_embedding=False):
Anthony Larcher's avatar
Anthony Larcher committed
148
149
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
150
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
151
        """
Anthony Larcher's avatar
Anthony Larcher committed
152
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
153
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
154
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
155
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
156

Anthony Larcher's avatar
Anthony Larcher committed
157
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
158
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
159
160
161
            self.activation = torch.nn.LeakyReLU(0.2)

            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
162

Anthony Larcher's avatar
xv    
Anthony Larcher committed
163
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
164
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
165
166
167
168
169
170
171
172
                ("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
173
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
174
175
                ("activation4", torch.nn.LeakyReLU(0.2)),
                ("norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
176
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
177
178
179
180
                ("activation5", torch.nn.LeakyReLU(0.2)),
                ("norm5", torch.nn.BatchNorm1d(1536))
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
181
182
            self.stat_pooling = MeanStdPooling()

Anthony Larcher's avatar
xv    
Anthony Larcher committed
183
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
184
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
185
186
            ]))

Anthony Larcher's avatar
xv    
Anthony Larcher committed
187
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
188
189
                ("activation6", torch.nn.LeakyReLU(0.2)),
                ("norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
190
                ("dropout6", torch.nn.Dropout(p=0.05)),
Anthony Larcher's avatar
Anthony Larcher committed
191
                ("linear7", torch.nn.Linear(512, 512)),
Anthony Larcher's avatar
Anthony Larcher committed
192
193
                ("activation7", torch.nn.LeakyReLU(0.2)),
                ("norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
194
                ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
Anthony Larcher's avatar
Anthony Larcher committed
195
196
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
197
198
199
200
            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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
        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
rawnet2    
Anthony Larcher committed
230
231
232
233
234
235
            self.preprocessor_weight_decay = 0.000
            self.sequence_network_weight_decay = 0.000
            self.stat_pooling_weight_decay = 0.000
            self.before_speaker_embedding_weight_decay = 0.00
            self.after_speaker_embedding_weight_decay = 0.00

Anthony Larcher's avatar
Anthony Larcher committed
236
237
        else:
            # Load Yaml configuration
Anthony Larcher's avatar
Anthony Larcher committed
238
            with open(model_archi, 'r') as fh:
Anthony Larcher's avatar
Anthony Larcher committed
239
240
                cfg = yaml.load(fh, Loader=yaml.FullLoader)

Anthony Larcher's avatar
Anthony Larcher committed
241
242
243
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
244
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
245
246
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
247
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
248
249
250
251
252
253
254
255
256
257
258
259
                        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
260
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
261
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
262
263
264
265
                    self.preprocessor = RawPreprocessor(nb_samp=48000,
                                                        in_channels=1,
                                                        filts=128,
                                                        first_conv=3)
Anthony Larcher's avatar
Anthony Larcher committed
266
                    self.feature_size = 128
Anthony Larcher's avatar
Anthony Larcher committed
267
268

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
269
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
270
            """
Anthony Larcher's avatar
Anthony Larcher committed
271
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
272
273
274
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
            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
291
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
292
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
293
294
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
295
296
297
298
299
300
301
302
                    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
303
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
304
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
305

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
306
307
308
309
310
311
312
313
314
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
            if cfg["stat_pooling"]["type"] == "GRU":
                self.stat_pooling = GruPooling(input_size=cfg["stat_pooling"]["input_size"],
                                               gru_node=cfg["stat_pooling"]["gru_node"],
                                               nb_gru_layer=cfg["stat_pooling"]["nb_gru_layer"])

Anthony Larcher's avatar
Anthony Larcher committed
315
316
317
            """
            Prepapre last part of the network (after pooling)
            """
Anthony Larcher's avatar
Anthony Larcher committed
318
319
            # Create sequential object for the second part of the network
            input_size = input_size * 2
Anthony Larcher's avatar
xv    
Anthony Larcher committed
320
321
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
322
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
323
324
                    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
325
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
326
327
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
328
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
329
330

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
331
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
332
333

                elif k.startswith('norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
334
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
335
336

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
337
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
338

Anthony Larcher's avatar
Anthony Larcher committed
339
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
340
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
341
342
343
344
345

            # 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
346
347
                    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
348
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
349
350
                        after_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
351
                        input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
352
353
354
355
356
357
358
359

                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
360
                    after_embedding_layers.append((k, torch.nn.Dropout(p=cfg["after_embedding"][k])))
Anthony Larcher's avatar
Anthony Larcher committed
361

Anthony Larcher's avatar
Anthony Larcher committed
362
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict(after_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
363
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
364

Anthony Larcher's avatar
Anthony Larcher committed
365
    def forward(self, x, is_eval=False):
366
367
368
        """

        :param x:
Anthony Larcher's avatar
Anthony Larcher committed
369
        :param is_eval:
370
371
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
372
373
374
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
375
        x = self.sequence_network(x)
376

Anthony Larcher's avatar
Anthony Larcher committed
377
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
378
379
380
381
        #mean = torch.mean(x, dim=2)
        #std = torch.std(x, dim=2)
        #x = torch.cat([mean, std], dim=1)
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
382
383
384
385

        x = self.before_speaker_embedding(x)
        if is_eval:
            return x
386

Anthony Larcher's avatar
Anthony Larcher committed
387
388
389
390
        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
391
392
        x = self.after_speaker_embedding(x)
        return x
Anthony Larcher's avatar
Anthony Larcher committed
393

Anthony Larcher's avatar
Anthony Larcher committed
394

Anthony Larcher's avatar
Anthony Larcher committed
395
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
396
           dataset_yaml,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
397
           epochs=100,
Anthony Larcher's avatar
Anthony Larcher committed
398
           lr=0.01,
Anthony Larcher's avatar
Anthony Larcher committed
399
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
400
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
401
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
402
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
403
           multi_gpu=True,
404
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
405
           opt='sgd',
Anthony Larcher's avatar
Anthony Larcher committed
406
407
           reset_parts=[],
           freeze_parts=[],
Anthony Larcher's avatar
Anthony Larcher committed
408
           num_thread=1):
409
410
    """

Anthony Larcher's avatar
Anthony Larcher committed
411
412
413
414
415
416
417
418
419
420
421
    :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:
422
423
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
424
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
425

Anthony Larcher's avatar
Anthony Larcher committed
426
427
    # Start from scratch
    if model_name is None:
Anthony Larcher's avatar
Anthony Larcher committed
428
        # Initialize a first model
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
429
        if model_yaml == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
430
            model = Xtractor(speaker_number, "xvector")
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
431
        elif model_yaml == "rawnet2":
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
432
            model = Xtractor(speaker_number, "rawnet2")
Anthony Larcher's avatar
Anthony Larcher committed
433
        else:
Anthony Larcher's avatar
Anthony Larcher committed
434
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
    # If we start from an existing model
    else:
        # Load the model
        logging.critical(f"*** Load model from = {model_name}")
        checkpoint = torch.load(model_name)
        model = Xtractor(speaker_number, model_yaml)

        """
        Here we remove all layers that we don't want to reload
        
        """
        pretrained_dict = checkpoint["model_state_dict"]
        for part in reset_parts:
            pretrained_dict = {k: v for k, v in pretrained_dict.items() if not k.startswith(part)}

Anthony Larcher's avatar
Anthony Larcher committed
450
        new_model_dict = model.state_dict()
Anthony Larcher's avatar
Anthony Larcher committed
451
452
453
454
455
456
457
        new_model_dict.update(pretrained_dict)
        model.load_state_dict(new_model_dict)

    # Freeze required layers
    for name, param in model.named_parameters():
        if name.split(".")[0] in freeze_parts:
            param.requires_grad = False
Anthony Larcher's avatar
Anthony Larcher committed
458

Anthony Larcher's avatar
Anthony Larcher committed
459
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
460
461
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
462
463
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
464
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
465

Anthony Larcher's avatar
Anthony Larcher committed
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
    if device.type == 'cuda':
        print(torch.cuda.get_device_name())
        print('Memory Usage:')
        print('Allocated:', round(torch.cuda.memory_allocated()/1024**3,1), 'GB')
        print('Cached:   ', round(torch.cuda.memory_cached()/1024**3,1), 'GB')

    test = torch.tensor(numpy.ones((128, 16000 * 5), dtype=numpy.float32))
    test.to(device)


    if device.type == 'cuda':
        print(torch.cuda.get_device_name())
        print('Memory Usage:')
        print('Allocated:', round(torch.cuda.memory_allocated()/1024**3,1), 'GB')
        print('Cached:   ', round(torch.cuda.memory_cached()/1024**3,1), 'GB')

Anthony Larcher's avatar
Anthony Larcher committed
482
    """
Anthony Larcher's avatar
Anthony Larcher committed
483
484
485
486
    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
487
    """
Anthony Larcher's avatar
Anthony Larcher committed
488
489
490
491
    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"])
492

Anthony Larcher's avatar
Anthony Larcher committed
493
    torch.manual_seed(dataset_params['seed'])
494
495
496
    training_set = SideSet(dataset_yaml, 
                           set_type="train", 
                           dataset_df=training_df, 
Anthony Larcher's avatar
Anthony Larcher committed
497
498
                           chunk_per_segment=dataset_params['train']['chunk_per_segment'], 
                           overlap=dataset_params['train']['overlap'])
Anthony Larcher's avatar
Anthony Larcher committed
499
500
501
    training_loader = DataLoader(training_set,
                                 batch_size=dataset_params["batch_size"],
                                 shuffle=True,
Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
502
                                 drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
503
                                 num_workers=num_thread)
504

Anthony Larcher's avatar
Anthony Larcher committed
505
506
507
    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
508
                                   drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
509
                                   num_workers=num_thread)
Anthony Larcher's avatar
Anthony Larcher committed
510
511
512
513

    """
    Set the training options
    """
Anthony Larcher's avatar
Anthony Larcher committed
514
515
    if opt == 'sgd':
        _optimizer = torch.optim.SGD
Anthony Larcher's avatar
Anthony Larcher committed
516
        _options = {'lr': lr, 'momentum': 0.9}
Anthony Larcher's avatar
Anthony Larcher committed
517
518
    elif opt == 'adam':
        _optimizer = torch.optim.Adam
Anthony Larcher's avatar
Anthony Larcher committed
519
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
520
521
    elif opt == 'rmsprop':
        _optimizer = torch.optim.RMSprop
Anthony Larcher's avatar
Anthony Larcher committed
522
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
523

Anthony Larcher's avatar
Anthony Larcher committed
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
    params = [
        {
            'params': [
                param for name, param in model.named_parameters() if 'bn' not in name
            ]
        },
        {
            'params': [
                param for name, param in model.named_parameters() if 'bn' in name
            ],
            'weight_decay': 0
        },
    ]

    optimizer = torch.optim.Adam(params,
                                 lr=0.001,
                                 weight_decay=0.0001,
                                 amsgrad=1)

    #if type(model) is Xtractor:
    #    optimizer = _optimizer([
    #        {'params': model.preprocessor.parameters(),
    #         'weight_decay': model.preprocessor_weight_decay},
    #        {'params': model.sequence_network.parameters(),
    #         'weight_decay': model.sequence_network_weight_decay},
    #        {'params': model.stat_pooling.parameters(),
    #         'weight_decay': model.stat_pooling_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}],
    #        **_options
    #    )
    #else:
    #    optimizer = _optimizer([
    #        {'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}],
    #        **_options
    #    )

Anthony Larcher's avatar
Anthony Larcher committed
568
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', verbose=True)
569

Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
570
    best_accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
571
    best_accuracy_epoch = 1
Anthony Larcher's avatar
Anthony Larcher committed
572
    for epoch in range(1, epochs + 1):
573
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
574
575
576
577
578
579
580
        model = train_epoch(model,
                            epoch,
                            training_loader,
                            optimizer,
                            dataset_params["log_interval"],
                            device=device,
                            clipping=clipping)
581
582

        # Add the cross validation here
Anthony Larcher's avatar
Anthony Larcher committed
583
        accuracy, val_loss = cross_validation(model, validation_loader, device=device)
584
585
586
587
        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
588
        print(f"Learning rate is {optimizer.param_groups[0]['lr']}")
589

Anthony Larcher's avatar
Anthony Larcher committed
590
591
592
593
        # remember best accuracy and save checkpoint
        is_best = accuracy > best_accuracy
        best_accuracy = max(accuracy, best_accuracy)

Anthony Larcher's avatar
Anthony Larcher committed
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
        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
610
611
612

        if is_best:
            best_accuracy_epoch = epoch
613

Anthony Larcher's avatar
Anthony Larcher committed
614
    logging.critical(f"Best accuracy {best_accuracy * 100.} obtained at epoch {best_accuracy_epoch}")
615

Anthony Larcher's avatar
Anthony Larcher committed
616
def train_epoch(model, epoch, training_loader, optimizer, log_interval, device, clipping=False):
617
618
619
620
    """

    :param model:
    :param epoch:
Anthony Larcher's avatar
Anthony Larcher committed
621
    :param training_loader:
622
    :param optimizer:
Anthony Larcher's avatar
Anthony Larcher committed
623
624
625
    :param log_interval:
    :param device:
    :param clipping:
626
627
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
628
    model.train()
Anthony Larcher's avatar
Anthony Larcher committed
629
    criterion = torch.nn.CrossEntropyLoss(reduction='mean')
630
631

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
632
    for batch_idx, (data, target) in enumerate(training_loader):
633
634
635
636
637
        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
638
639
        if clipping:
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
640
641
642
        optimizer.step()
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()

Anthony Larcher's avatar
Anthony Larcher committed
643
        if batch_idx % log_interval == 0:
Anthony Larcher's avatar
Anthony Larcher committed
644
            batch_size = target.shape[0]
645
            logging.critical('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
Anthony Larcher's avatar
Anthony Larcher committed
646
                epoch, batch_idx + 1, training_loader.__len__(),
Anthony Larcher's avatar
Anthony Larcher committed
647
648
                100. * batch_idx / training_loader.__len__(), loss.item(),
                100.0 * accuracy.item() / ((batch_idx + 1) * batch_size)))
649
650
651
    return model


Anthony Larcher's avatar
Anthony Larcher committed
652
def cross_validation(model, validation_loader, device):
653
654
655
    """

    :param model:
Anthony Larcher's avatar
Anthony Larcher committed
656
657
    :param validation_loader:
    :param device:
658
659
660
661
662
    :return:
    """
    model.eval()

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
663
    loss = 0.0
664
    criterion = torch.nn.CrossEntropyLoss()
Anthony Larcher's avatar
Anthony Larcher committed
665
666
667
668
669
670
    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()
671

Anthony Larcher's avatar
Anthony Larcher committed
672
673
            loss += criterion(output, target.to(device))
    
Anthony Larcher's avatar
Anthony Larcher committed
674
675
676
677
    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
678
679
680
681
682
683
684
685
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
686

Anthony Larcher's avatar
Anthony Larcher committed
687
    if isinstance(idmap_name, IdMap):
688
689
690
691
        idmap = idmap_name
    else:
        idmap = IdMap(idmap_name)

Anthony Larcher's avatar
Anthony Larcher committed
692
    # Create dataset to load the data
Anthony Larcher's avatar
Anthony Larcher committed
693
694
695
696
    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
697
698

    # Load the model
699
700
701
702
703
704
705
    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
706
707
    model.eval()
    model.to(device)
708

Anthony Larcher's avatar
Anthony Larcher committed
709
710
711
    # 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
712
    
Anthony Larcher's avatar
Anthony Larcher committed
713
    # Create the StatServer
Anthony Larcher's avatar
Anthony Larcher committed
714
    embeddings = StatServer()
Anthony Larcher's avatar
Anthony Larcher committed
715
716
717
718
719
720
    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
721

Anthony Larcher's avatar
Anthony Larcher committed
722
723
    # Process the data
    with torch.no_grad():
Anthony Larcher's avatar
Anthony Larcher committed
724
        for idx in tqdm.tqdm(range(len(dataset))):
Anthony Larcher's avatar
Anthony Larcher committed
725
            data, mod, seg, start, stop = dataset[idx]
Anthony Larcher's avatar
Anthony Larcher committed
726
            vec = model(data[None, :, :].to(device), is_eval=True)
Anthony Larcher's avatar
Anthony Larcher committed
727
728
729
730
731
732
            #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
733
734
735
736

    return embeddings


Anthony Larcher's avatar
Anthony Larcher committed
737
738
739
740
741
742
743
744
745
746
747
748
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):

749
750
751
752
753
754
755
756
757
758
759

    # 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):
Anthony Larcher's avatar
Anthony Larcher committed
760

761
762
763
        # 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
764
                                    int(sample_rate * (window_length - overlap)))
765
766

        # Create a numpy array to store the current x-vectors
Anthony Larcher's avatar
Anthony Larcher committed
767
        model_names.append(numpy.array([mod + f"_{ii}" for ii in range(len(chunk_starts))]).astype("U"))
768
769
770
771
772
773
774
        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
775
776
777
778
    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)
779
    assert sliding_idmap.validate()
Anthony Larcher's avatar
Anthony Larcher committed
780

Anthony Larcher's avatar
Anthony Larcher committed
781
782
783
784
785
786
    embeddings = extract_embeddings(sliding_idmap,
                                 speaker_number,
                                 model_filename,
                                 model_yaml,
                                 data_root_name,
                                 device,
Anthony Larcher's avatar
Anthony Larcher committed
787
788
                                 file_extension=file_extension,
                                 transform_pipeline=transform_pipeline)
Anthony Larcher's avatar
Anthony Larcher committed
789
790

    return embeddings
Anthony Larcher's avatar
Anthony Larcher committed
791
792