xvector.py 36.7 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
30

import pdb
import traceback
Anthony Larcher's avatar
Anthony Larcher committed
31
import logging
Anthony Larcher's avatar
Anthony Larcher committed
32
import matplotlib.pyplot as plt
Anthony Larcher's avatar
Anthony Larcher committed
33
import numpy
Anthony Larcher's avatar
Anthony Larcher committed
34
import pandas
Anthony Larcher's avatar
minor    
Anthony Larcher committed
35
import pickle
Anthony Larcher's avatar
Anthony Larcher committed
36
import shutil
37
import time
Anthony Larcher's avatar
Anthony Larcher committed
38
import torch
Anthony Larcher's avatar
Anthony Larcher committed
39
import torch.optim as optim
Anthony Larcher's avatar
Anthony Larcher committed
40
41
import yaml

Anthony Larcher's avatar
Anthony Larcher committed
42
from torchvision import transforms
Anthony Larcher's avatar
Anthony Larcher committed
43
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
44
from .xsets import SideSet
Anthony Larcher's avatar
Anthony Larcher committed
45
from .xsets import IdMapSet
Anthony Larcher's avatar
Anthony Larcher committed
46
from .res_net import RawPreprocessor, ResBlockWFMS
Anthony Larcher's avatar
Anthony Larcher committed
47
48
from ..bosaris import IdMap
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
49
from torch.utils.data import DataLoader
Anthony Larcher's avatar
Anthony Larcher committed
50
from sklearn.model_selection import train_test_split
Anthony Larcher's avatar
Anthony Larcher committed
51
from .sincnet import SincNet
Anthony Larcher's avatar
Anthony Larcher committed
52
#from torch.utils.tensorboard import SummaryWriter
Anthony Larcher's avatar
Anthony Larcher committed
53
from .loss import ArcLinear
Anthony Larcher's avatar
Anthony Larcher committed
54

Anthony Larcher's avatar
Anthony Larcher committed
55
import tqdm
Anthony Larcher's avatar
Anthony Larcher committed
56

Anthony Larcher's avatar
Anthony Larcher committed
57
58
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
59
__copyright__ = "Copyright 2015-2020 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
60
61
62
63
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
64
65


Anthony Larcher's avatar
Anthony Larcher committed
66
67
logging.basicConfig(format='%(asctime)s %(message)s')

Anthony Larcher's avatar
Anthony Larcher committed
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
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
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157

# Make PyTorch Deterministic
torch.manual_seed(0)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
numpy.random.seed(0)


class GuruMeditation (torch.autograd.detect_anomaly):
    
    def __init__(self):
        super(GuruMeditation, self).__init__()

    def __enter__(self):
        super(GuruMeditation, self).__enter__()
        return self

    def __exit__(self, type, value, trace):
        super(GuruMeditation, self).__exit__()
        if isinstance(value, RuntimeError):
            traceback.print_tb(trace)
            halt(str(value))

    def halt(msg):
        print (msg)
        pdb.set_trace()









def select_n_random(data, labels, n=100):
    '''
    Selects n random datapoints and their corresponding labels from a dataset
    '''
    assert len(data) == len(labels)

    perm = torch.randperm(len(data))
    return data[perm][:n], labels[perm][:n]


def matplotlib_imshow(img, one_channel=False):
    if one_channel:
        img = img.mean(dim=0)
    img = img / 2 + 0.5     # unnormalize
    npimg = img.cpu().numpy()
    if one_channel:
        plt.imshow(npimg, cmap="Greys")
    else:
        plt.imshow(np.transpose(npimg, (1, 2, 0)))

def speech_to_probs(model, speech):
    '''
    Generates predictions and corresponding probabilities from a trained
    network and a list of images
    '''
    output = model(speech)
    # convert output probabilities to predicted class
    _, preds_tensor = torch.max(output, 1)
    preds = numpy.squeeze(preds_tensor.cpu().numpy())
    return preds, [torch.nn.functional.softmax(el, dim=0)[i].item() for i, el in zip(preds, output)]


def plot_classes_preds(model, speech, labels):
    '''
    Generates matplotlib Figure using a trained network, along with images
    and labels from a batch, that shows the network's top prediction along
    with its probability, alongside the actual label, coloring this
    information based on whether the prediction was correct or not.
    Uses the "speech_to_probs" function.
    '''
    preds, probs = speech_to_probs(model, speech)
    # plot the images in the batch, along with predicted and true labels
    fig = plt.figure(figsize=(12, 48))
    for idx in numpy.arange(4):
         ax = fig.add_subplot(1, 4, idx+1, xticks=[], yticks=[])
         #matplotlib_imshow(speech[idx], one_channel=True)
         ax.set_title("{0}, {1:.1f}%\n(label: {2})".format(
                                                           preds[idx],
                                                           probs[idx] * 100.0,
                                                           labels[idx]),
                                                           color=("green" if preds[idx]==labels[idx].item() else "red"))
    return fig



158
def get_lr(optimizer):
Anthony Larcher's avatar
Anthony Larcher committed
159
160
161
162
163
    """

    :param optimizer:
    :return:
    """
164
165
166
167
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
168
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
169
170
171
172
173
174
175
176
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
177
178
179
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
180

Anthony Larcher's avatar
Anthony Larcher committed
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
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
201

Anthony Larcher's avatar
Anthony Larcher committed
202

Anthony Larcher's avatar
Anthony Larcher committed
203
204
205
206
207
208
209
210
211
212
213
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
214
        super(GruPooling, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
        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
237

Anthony Larcher's avatar
Anthony Larcher committed
238
class Xtractor(torch.nn.Module):
239
240
241
    """
    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
242

Anthony Larcher's avatar
Anthony Larcher committed
243
244
245
246
247
248
249
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
                 loss="cce",
                 norm_embedding=False,
                 aam_margin=0.5,
                 aam_s=0.5):
Anthony Larcher's avatar
Anthony Larcher committed
250
251
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
252
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
253
        """
Anthony Larcher's avatar
Anthony Larcher committed
254
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
255
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
256
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
257
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
258

Anthony Larcher's avatar
Anthony Larcher committed
259
260
261
262
        self.loss = loss
        if loss not in ["cce", 'aam']:
            raise NotImplementedError(f"The valid loss are for now cce and aam ")

Anthony Larcher's avatar
Anthony Larcher committed
263
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
264
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
265
266
267
            self.activation = torch.nn.LeakyReLU(0.2)

            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
268

Anthony Larcher's avatar
xv    
Anthony Larcher committed
269
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
270
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
271
272
273
274
275
276
277
278
                ("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
279
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
280
281
                ("activation4", torch.nn.LeakyReLU(0.2)),
                ("norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
282
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
283
284
285
286
                ("activation5", torch.nn.LeakyReLU(0.2)),
                ("norm5", torch.nn.BatchNorm1d(1536))
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
287
288
            self.stat_pooling = MeanStdPooling()

Anthony Larcher's avatar
xv    
Anthony Larcher committed
289
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
290
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
291
292
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
293
294
295
296
297
298
299
300
301
302
303
304
305
306
            if self.loss == "aam":
                self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
                  ("arclinear8", ArcLinear(512, int(self.speaker_number), margin=aam_margin, s=aam_s))
                ]))
            elif self.loss == "cce":
                self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
                    ("activation6", torch.nn.LeakyReLU(0.2)),
                    ("norm6", torch.nn.BatchNorm1d(512)),
                    ("dropout6", torch.nn.Dropout(p=0.05)),
                    ("linear7", torch.nn.Linear(512, 512)),
                    ("activation7", torch.nn.LeakyReLU(0.2)),
                    ("norm7", torch.nn.BatchNorm1d(512)),
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
307

Anthony Larcher's avatar
Anthony Larcher committed
308
309
310
311
            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
312
313
314
315
        elif model_archi == "rawnet2":
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
316
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
317
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
318
319
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336

            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)

Anthony Larcher's avatar
Anthony Larcher committed
337
338
339
340
341
342
343
344
345
            if self.loss == "aam":
                if loss == 'aam':
                    self.after_speaker_embedding = ArcLinear(1024,
                                                             int(self.speaker_number),
                                                             margin=aam_margin, s=aam_s)
            elif self.loss == "cce"
                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
346

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
347
348
349
350
351
352
            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
353
354
        else:
            # Load Yaml configuration
Anthony Larcher's avatar
Anthony Larcher committed
355
            with open(model_archi, 'r') as fh:
Anthony Larcher's avatar
Anthony Larcher committed
356
357
                cfg = yaml.load(fh, Loader=yaml.FullLoader)

Anthony Larcher's avatar
Anthony Larcher committed
358
359
360
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
361
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
362
363
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
364
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
365
366
367
368
369
370
371
372
373
374
375
376
                        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
377
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
378
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
379
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
380
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
381
382
383
384
385
386
                                                        out_channels=cfg["feature_size"],
                                                        kernel_size=cfg['preprocessor']["kernel_size"],
                                                        stride=cfg['preprocessor']["stride"],
                                                        padding=cfg['preprocessor']["padding"],
                                                        dilation=cfg['preprocessor']["dilation"])
                    self.feature_size = cfg["feature_size"]
Anthony Larcher's avatar
Anthony Larcher committed
387
388

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
389
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
390
            """
Anthony Larcher's avatar
Anthony Larcher committed
391
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
392
393
394
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
            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():
Anthony Larcher's avatar
Anthony Larcher committed
410
411
412
413
414
415
                if k.startswith("lin"):
                    segmental_layers.append((k, torch.nn.Linear(input_size,
                                                                cfg["segmental"][k]["output"])))
                    input_size = cfg["segmental"][k]["output"]

                elif k.startswith("conv"):
Anthony Larcher's avatar
Anthony Larcher committed
416
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
417
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
418
419
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
420
421
422
423
424
425
426
427
                    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
428
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
429
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
430

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
431
432
433
434
435
436
437
438
439
            """
            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
440
441
442
            """
            Prepapre last part of the network (after pooling)
            """
Anthony Larcher's avatar
Anthony Larcher committed
443
444
            # Create sequential object for the second part of the network
            input_size = input_size * 2
Anthony Larcher's avatar
xv    
Anthony Larcher committed
445
446
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
447
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
448
449
                    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
450
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
451
452
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
453
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
454
455

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
456
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
457
458

                elif k.startswith('norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
459
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
460
461

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
462
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
463

Anthony Larcher's avatar
Anthony Larcher committed
464
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
465
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
466
467
468
469
470

            # 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
471
472
                    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
473
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
474
475
                        after_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
476
                        input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
477

Anthony Larcher's avatar
Anthony Larcher committed
478
479
480
481
482
483
484
485
486
487
488
489
                elif k.startswith('arc'):
                    if cfg["after_embedding"][k]["output"] == "speaker_number":
                        after_embedding_layers.append(
                            (k, ArcLinear(input_size, self.speaker_number, margin=aam_margin, s=aam_s)))
                    else:
                        after_embedding_layers.append(
                            (k, ArcLinear(input_size,
                                          self.speaker_number,
                                          margin=aam_margin,
                                          s=aam_s)))
                        input_size = self.speaker_number

Anthony Larcher's avatar
xv    
Anthony Larcher committed
490
491
492
493
494
495
496
                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
497
                    after_embedding_layers.append((k, torch.nn.Dropout(p=cfg["after_embedding"][k])))
Anthony Larcher's avatar
Anthony Larcher committed
498

Anthony Larcher's avatar
Anthony Larcher committed
499
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict(after_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
500
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
501

Anthony Larcher's avatar
Anthony Larcher committed
502

Anthony Larcher's avatar
Anthony Larcher committed
503
    def forward(self, x, is_eval=False, target=None):
504
505
506
        """

        :param x:
Anthony Larcher's avatar
Anthony Larcher committed
507
        :param is_eval:
508
509
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
510
511
        if self.preprocessor is not None:
            x = self.preprocessor(x)
Anthony Larcher's avatar
Anthony Larcher committed
512
            print("go through preprocessor")
Anthony Larcher's avatar
Anthony Larcher committed
513

Anthony Larcher's avatar
Anthony Larcher committed
514
        x = self.sequence_network(x)
515

Anthony Larcher's avatar
Anthony Larcher committed
516
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
517
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
518
519
520
521

        x = self.before_speaker_embedding(x)
        if is_eval:
            return x
522

Anthony Larcher's avatar
Anthony Larcher committed
523
524
525
526
        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
527
528
529
530
531
532
533
534
        if self.loss == "cce":
            x = self.after_speaker_embedding(x)
        elif self.loss == "aam":
            if not is_eval:
                x = self.after_speaker_embedding(x,target=target)
            else:
                x = self.after_speaker_embedding(x, target=None)

Anthony Larcher's avatar
Anthony Larcher committed
535
        return x
Anthony Larcher's avatar
Anthony Larcher committed
536

Anthony Larcher's avatar
Anthony Larcher committed
537

Anthony Larcher's avatar
Anthony Larcher committed
538
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
539
           dataset_yaml,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
540
           epochs=100,
Anthony Larcher's avatar
Anthony Larcher committed
541
           lr=0.01,
Anthony Larcher's avatar
Anthony Larcher committed
542
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
543
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
544
545
546
547
           loss="cce",
           aam_margin=0.5,
           aam_s=30,
           patience=10,
Anthony Larcher's avatar
Anthony Larcher committed
548
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
549
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
550
           multi_gpu=True,
551
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
552
           opt='sgd',
Anthony Larcher's avatar
Anthony Larcher committed
553
554
           reset_parts=[],
           freeze_parts=[],
Anthony Larcher's avatar
Anthony Larcher committed
555
           num_thread=1):
556
557
    """

Anthony Larcher's avatar
Anthony Larcher committed
558
559
560
561
562
563
    :param speaker_number:
    :param dataset_yaml:
    :param epochs:
    :param lr:
    :param model_yaml:
    :param model_name:
Anthony Larcher's avatar
Anthony Larcher committed
564
565
566
567
    :param loss:
    :param aam_margin:
    :param aam_s:
    :param patience:
Anthony Larcher's avatar
Anthony Larcher committed
568
569
570
571
    :param tmp_model_name:
    :param best_model_name:
    :param multi_gpu:
    :param clipping:
Anthony Larcher's avatar
Anthony Larcher committed
572
573
574
    :param opt:
    :param reset_parts:
    :param freeze_parts:
Anthony Larcher's avatar
Anthony Larcher committed
575
    :param num_thread:
576
577
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
578
579
580
581
    # Add for tensorboard
    # Writer will output to ./runs/ directory by default
    #writer = SummaryWriter("runs/xvectors_experiments_2")
    writer = None
Anthony Larcher's avatar
Anthony Larcher committed
582

Anthony Larcher's avatar
Anthony Larcher committed
583
    logging.critical(f"Start process at {time.strftime('%H:%M:%S', time.localtime())}")
584

Anthony Larcher's avatar
Anthony Larcher committed
585
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
586

Anthony Larcher's avatar
Anthony Larcher committed
587
588
    # Start from scratch
    if model_name is None:
Anthony Larcher's avatar
Anthony Larcher committed
589
        # Initialize a first model
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
590
        if model_yaml == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
591
            model = Xtractor(speaker_number, "xvector")
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
592
        elif model_yaml == "rawnet2":
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
593
            model = Xtractor(speaker_number, "rawnet2")
Anthony Larcher's avatar
Anthony Larcher committed
594
        else:
Anthony Larcher's avatar
Anthony Larcher committed
595
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
    # 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
611
        new_model_dict = model.state_dict()
Anthony Larcher's avatar
Anthony Larcher committed
612
613
614
615
616
617
618
        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
619

Anthony Larcher's avatar
Anthony Larcher committed
620

Anthony Larcher's avatar
Anthony Larcher committed
621
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
622
623
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
624
625
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
626
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
627
628

    """
Anthony Larcher's avatar
Anthony Larcher committed
629
630
631
632
    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
633
    """
Anthony Larcher's avatar
Anthony Larcher committed
634
635
636
637
    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"])
638

Anthony Larcher's avatar
Anthony Larcher committed
639
    torch.manual_seed(dataset_params['seed'])
640
641
642
    training_set = SideSet(dataset_yaml, 
                           set_type="train", 
                           dataset_df=training_df, 
Anthony Larcher's avatar
Anthony Larcher committed
643
644
                           chunk_per_segment=dataset_params['train']['chunk_per_segment'], 
                           overlap=dataset_params['train']['overlap'])
Anthony Larcher's avatar
Anthony Larcher committed
645
646
647
    training_loader = DataLoader(training_set,
                                 batch_size=dataset_params["batch_size"],
                                 shuffle=True,
Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
648
                                 drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
649
                                 pin_memory=True,
Anthony Larcher's avatar
Anthony Larcher committed
650
                                 num_workers=num_thread)
651

Anthony Larcher's avatar
Anthony Larcher committed
652
653
654
    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
655
                                   drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
656
                                   pin_memory=True,
Anthony Larcher's avatar
Anthony Larcher committed
657
                                   num_workers=num_thread)
Anthony Larcher's avatar
Anthony Larcher committed
658

Anthony Larcher's avatar
Anthony Larcher committed
659
660
661
662
663
664
    # Add for TensorBoard
    #dataiter = iter(training_loader)
    #data, labels = dataiter.next()
    #writer.add_graph(model, data)


Anthony Larcher's avatar
Anthony Larcher committed
665
666
667
    """
    Set the training options
    """
Anthony Larcher's avatar
Anthony Larcher committed
668
    if opt == 'adam':
Anthony Larcher's avatar
Anthony Larcher committed
669
        _optimizer = torch.optim.Adam
Anthony Larcher's avatar
Anthony Larcher committed
670
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
671
672
    elif opt == 'rmsprop':
        _optimizer = torch.optim.RMSprop
Anthony Larcher's avatar
Anthony Larcher committed
673
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
674
675
676
    else: # opt == 'sgd'
        _optimizer = torch.optim.SGD
        _options = {'lr': lr, 'momentum': 0.9}
Anthony Larcher's avatar
Anthony Larcher committed
677

Anthony Larcher's avatar
Anthony Larcher committed
678
679
680
681
682
683
684
685
686
687
688
689
690
691
    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
        },
    ]

Anthony Larcher's avatar
Anthony Larcher committed
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
    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
        )

    #optimizer = torch.optim.SGD(params,
    #                            lr=lr,
    #                            momentum=0.9,
    #                            weight_decay=0.0005)
    #print(f"Learning rate = {lr}")
Anthony Larcher's avatar
Anthony Larcher committed
722

Anthony Larcher's avatar
Anthony Larcher committed
723
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', verbose=True)
724

Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
725
    best_accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
726
    best_accuracy_epoch = 1
Anthony Larcher's avatar
Anthony Larcher committed
727
    curr_patience = patience
Anthony Larcher's avatar
Anthony Larcher committed
728
    for epoch in range(1, epochs + 1):
729
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
730
731
732
        if curr_patience == 0:
            print(f"Stopping at epoch {epoch} for cause of patience")
            break
Anthony Larcher's avatar
Anthony Larcher committed
733
734
735
736
737
738
        model = train_epoch(model,
                            epoch,
                            training_loader,
                            optimizer,
                            dataset_params["log_interval"],
                            device=device,
Anthony Larcher's avatar
Anthony Larcher committed
739
740
                            clipping=clipping,
                            tb_writer=writer)
741
742

        # Add the cross validation here
Anthony Larcher's avatar
Anthony Larcher committed
743
        accuracy, val_loss = cross_validation(model, validation_loader, device=device)
Anthony Larcher's avatar
Anthony Larcher committed
744
        logging.critical(f"***{time.strftime('%H:%M:%S', time.localtime())} Cross validation accuracy = {accuracy} %")
745
746
747
748

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

Anthony Larcher's avatar
Anthony Larcher committed
749
750
751
752
        # remember best accuracy and save checkpoint
        is_best = accuracy > best_accuracy
        best_accuracy = max(accuracy, best_accuracy)

Anthony Larcher's avatar
Anthony Larcher committed
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
        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
769
770
771

        if is_best:
            best_accuracy_epoch = epoch
Anthony Larcher's avatar
Anthony Larcher committed
772
773
774
            curr_patience = patience
        else:
            curr_patience -= 1
Anthony Larcher's avatar
Anthony Larcher committed
775
    #writer.close()
776

777
778
779
    for ii in range(torch.cuda.device_count()):
        print(torch.cuda.memory_summary(ii))

Anthony Larcher's avatar
Anthony Larcher committed
780
    logging.critical(f"Best accuracy {best_accuracy * 100.} obtained at epoch {best_accuracy_epoch}")
781

Anthony Larcher's avatar
Anthony Larcher committed
782
def train_epoch(model, epoch, training_loader, optimizer, log_interval, device, clipping=False, tb_writer=None):
783
784
785
786
    """

    :param model:
    :param epoch:
Anthony Larcher's avatar
Anthony Larcher committed
787
    :param training_loader:
788
    :param optimizer:
Anthony Larcher's avatar
Anthony Larcher committed
789
790
791
    :param log_interval:
    :param device:
    :param clipping:
792
793
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
794
    model.train()
Anthony Larcher's avatar
Anthony Larcher committed
795
    criterion = torch.nn.CrossEntropyLoss(reduction='mean')
796
797

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
798
    running_loss = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
799
    for batch_idx, (data, target) in enumerate(training_loader):
800
801
802
        target = target.squeeze()
        optimizer.zero_grad()
        output = model(data.to(device))
Anthony Larcher's avatar
Anthony Larcher committed
803
        #with GuruMeditation():
804
        loss = criterion(output, target.to(device))
Anthony Larcher's avatar
Anthony Larcher committed
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
        if not torch.isnan(loss):
            loss.backward()
            if clipping:
                torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
            running_loss += loss.item()
            optimizer.step()

            running_loss += loss.item()
            accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()

            if batch_idx % log_interval == 0:
                batch_size = target.shape[0]
                logging.critical('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
                    epoch, batch_idx + 1, training_loader.__len__(),
                    100. * batch_idx / training_loader.__len__(), loss.item(),
                    100.0 * accuracy.item() / ((batch_idx + 1) * batch_size)))
                #tb_writer.add_scalar('training loss',
                #                     running_loss / log_interval,
                #                     epoch * len(training_loader) + batch_idx)
                #tb_writer.add_scalar('training_accuracy',
                #                      100.0 * accuracy.item() / ((batch_idx + 1) * batch_size),
                #                      epoch * len(training_loader) + batch_idx)

                # ...log a Matplotlib Figure showing the model's predictions on a
                # random mini-batch
                #tb_writer.add_figure('predictions vs. actuals',
                #                     plot_classes_preds(model, data.to(device), target.to(device)),
                #                     global_step=epoch * len(training_loader) + batch_idx)

834
835
836
837
838
839
840
841
842
843
844
845
846
        else:
            save_checkpoint({
                             'epoch': epoch,
                             'model_state_dict': model.state_dict(),
                             'optimizer_state_dict': optimizer.state_dict(),
                             'accuracy': 0.0,
                             'scheduler': 0.0
                             }, False, filename="model_loss_NAN.pt", best_filename='toto.pt')
            with open("batch_loss_NAN.pkl", "wb") as fh:
                pickle.dump(data.cpu(), fh)
            import sys
            sys.exit()
        running_loss = 0.0
847
848
849
    return model


Anthony Larcher's avatar
Anthony Larcher committed
850
def cross_validation(model, validation_loader, device):
851
852
853
    """

    :param model:
Anthony Larcher's avatar
Anthony Larcher committed
854
855
    :param validation_loader:
    :param device:
856
857
858
859
860
    :return:
    """
    model.eval()

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
861
    loss = 0.0
862
    criterion = torch.nn.CrossEntropyLoss()
Anthony Larcher's avatar
Anthony Larcher committed
863
864
865
866
867
868
    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()
869

Anthony Larcher's avatar
Anthony Larcher committed
870
871
            loss += criterion(output, target.to(device))
    
Anthony Larcher's avatar
Anthony Larcher committed
872
873
874
875
    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
876
877
878
879
880
881
882
883
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
884

Anthony Larcher's avatar
Anthony Larcher committed
885
    if isinstance(idmap_name, IdMap):
886
887
888
889
        idmap = idmap_name
    else:
        idmap = IdMap(idmap_name)

Anthony Larcher's avatar
Anthony Larcher committed
890
    # Create dataset to load the data
Anthony Larcher's avatar
Anthony Larcher committed
891
892
893
894
    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
895
896

    # Load the model
897
898
899
900
901
902
903
    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
904
905
    model.eval()
    model.to(device)
906

Anthony Larcher's avatar
Anthony Larcher committed
907
908
909
    # 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
910
    
Anthony Larcher's avatar
Anthony Larcher committed
911
    # Create the StatServer
Anthony Larcher's avatar
Anthony Larcher committed
912
    embeddings = StatServer()
Anthony Larcher's avatar
Anthony Larcher committed
913
914
915
916
917
918
    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
919

Anthony Larcher's avatar
Anthony Larcher committed
920
921
    # Process the data
    with torch.no_grad():
Anthony Larcher's avatar
Anthony Larcher committed
922
        for idx in tqdm.tqdm(range(len(dataset))):
Anthony Larcher's avatar
Anthony Larcher committed
923
            data, mod, seg, start, stop = dataset[idx]
Anthony Larcher's avatar
Anthony Larcher committed
924
            vec = model(data[None, :, :].to(device), is_eval=True)
Anthony Larcher's avatar
Anthony Larcher committed
925
926
927
928
929
930
            #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
931
932
933
934

    return embeddings


Anthony Larcher's avatar
Anthony Larcher committed
935
936
937
938
939
940
941
942
943
944
945
946
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):

947
948
949
950
951
952
953
954
955
956
957

    # 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
958

959
960
961
        # 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
962
                                    int(sample_rate * (window_length - overlap)))
963
964

        # Create a numpy array to store the current x-vectors
Anthony Larcher's avatar
Anthony Larcher committed
965
        model_names.append(numpy.array([mod + f"_{ii}" for ii in range(len(chunk_starts))]).astype("U"))
966
967
968
969
970
971
972
        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
973
974
975
976
    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)
977
    assert sliding_idmap.validate()
Anthony Larcher's avatar
Anthony Larcher committed
978

Anthony Larcher's avatar
Anthony Larcher committed
979
980
981
982
983
984
    embeddings = extract_embeddings(sliding_idmap,
                                 speaker_number,
                                 model_filename,
                                 model_yaml,
                                 data_root_name,
                                 device,
Anthony Larcher's avatar
Anthony Larcher committed
985
986
                                 file_extension=file_extension,
                                 transform_pipeline=transform_pipeline)
Anthony Larcher's avatar
Anthony Larcher committed
987
988

    return embeddings
Anthony Larcher's avatar
Anthony Larcher committed
989