xvector.py 37.4 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
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
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
246
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
247
248
249
                 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
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
260
261
262
263
264
265

            if loss not in ["cce", 'aam']:
                raise NotImplementedError(f"The valid loss are for now cce and aam ")
            else:
                self.loss = loss

Anthony Larcher's avatar
Anthony Larcher committed
266
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
267
268
269
            self.activation = torch.nn.LeakyReLU(0.2)

            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
270

Anthony Larcher's avatar
xv    
Anthony Larcher committed
271
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
272
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
273
274
275
276
277
278
279
280
                ("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
281
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
282
283
                ("activation4", torch.nn.LeakyReLU(0.2)),
                ("norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
284
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
285
286
287
288
                ("activation5", torch.nn.LeakyReLU(0.2)),
                ("norm5", torch.nn.BatchNorm1d(1536))
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
289
290
            self.stat_pooling = MeanStdPooling()

Anthony Larcher's avatar
xv    
Anthony Larcher committed
291
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
292
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
293
294
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
295
296
            if self.loss == "aam":
                self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
297
                  ("arclinear", ArcLinear(512, int(self.speaker_number), margin=aam_margin, s=aam_s))
Anthony Larcher's avatar
Anthony Larcher committed
298
299
300
301
302
303
304
305
306
307
308
                ]))
            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
309

Anthony Larcher's avatar
Anthony Larcher committed
310
311
312
313
            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
314
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
315
316
317
318
319
320

            if loss not in ["cce", 'aam']:
                raise NotImplementedError(f"The valid loss are for now cce and aam ")
            else:
                self.loss = loss

Anthony Larcher's avatar
Anthony Larcher committed
321
322
323
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
324
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
325
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
326
327
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344

            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
345
346
347
348
349
350
351
352
353
            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
354

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
355
356
357
358
359
360
            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
361
362
        else:
            # Load Yaml configuration
Anthony Larcher's avatar
Anthony Larcher committed
363
            with open(model_archi, 'r') as fh:
Anthony Larcher's avatar
Anthony Larcher committed
364
365
                cfg = yaml.load(fh, Loader=yaml.FullLoader)

Anthony Larcher's avatar
Anthony Larcher committed
366
367
368
369
370
            self.loss = cfg["loss"]
            if self.loss == "aam":
                self.aam_margin = cfg["aam_margin"]
                self.aam_s = cfg["aam_s"]

Anthony Larcher's avatar
Anthony Larcher committed
371
372
373
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
374
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
375
376
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
377
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
378
379
380
381
382
383
384
385
386
387
388
389
                        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
390
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
391
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
392
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
393
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
394
395
396
397
398
399
                                                        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
400
401

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
402
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
403
            """
Anthony Larcher's avatar
Anthony Larcher committed
404
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
405
406
407
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
            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
423
424
425
426
427
428
                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
429
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
430
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
431
432
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
433
434
435
436
437
                    input_size = cfg["segmental"][k]["output_channels"]

                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
438
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
Anthony Larcher committed
439
440
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
441
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
442
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
443

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
444
445
446
447
448
449
450
451
452
            """
            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
453
            """
Anthony Larcher's avatar
Anthony Larcher committed
454
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
455
            """
Anthony Larcher's avatar
Anthony Larcher committed
456
457
            # Create sequential object for the second part of the network
            input_size = input_size * 2
Anthony Larcher's avatar
xv    
Anthony Larcher committed
458
459
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
460
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
461
462
                    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
463
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
464
465
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
466
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
467
468

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
469
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
470

Anthony Larcher's avatar
Anthony Larcher committed
471
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
472
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
473
474

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
475
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
476

Anthony Larcher's avatar
Anthony Larcher committed
477
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
478
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
479
480
481
482
483

            # 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
484
485
                    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
486
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
487
488
                        after_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
489
                        input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
490

Anthony Larcher's avatar
Anthony Larcher committed
491
                elif k.startswith('arc'):
Anthony Larcher's avatar
Anthony Larcher committed
492
493
494
495
                    after_embedding_layers.append((k, ArcLinear(input_size,
                                                                self.speaker_number,
                                                                margin=self.aam_margin,
                                                                s=self.aam_s)))
Anthony Larcher's avatar
Anthony Larcher committed
496

Anthony Larcher's avatar
xv    
Anthony Larcher committed
497
498
499
                elif k.startswith("activation"):
                    after_embedding_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
500
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
501
502
503
                    after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
504
                    after_embedding_layers.append((k, torch.nn.Dropout(p=cfg["after_embedding"][k])))
Anthony Larcher's avatar
Anthony Larcher committed
505

Anthony Larcher's avatar
Anthony Larcher committed
506
            self.after_speaker_embedding = torch.nn.Sequential(OrderedDict(after_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
507
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
508

Anthony Larcher's avatar
Anthony Larcher committed
509

Anthony Larcher's avatar
Anthony Larcher committed
510
    def forward(self, x, is_eval=False, target=None):
511
512
513
        """

        :param x:
Anthony Larcher's avatar
Anthony Larcher committed
514
        :param is_eval:
515
516
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
517
518
        if self.preprocessor is not None:
            x = self.preprocessor(x)
Anthony Larcher's avatar
Anthony Larcher committed
519
            print("go through preprocessor")
Anthony Larcher's avatar
Anthony Larcher committed
520

Anthony Larcher's avatar
Anthony Larcher committed
521
        x = self.sequence_network(x)
522

Anthony Larcher's avatar
Anthony Larcher committed
523
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
524
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
525
526

        x = self.before_speaker_embedding(x)
527

Anthony Larcher's avatar
Anthony Larcher committed
528
529
530
531
        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
532
533
534
        if is_eval:
            return x

Anthony Larcher's avatar
Anthony Larcher committed
535
536
        if self.loss == "cce":
            x = self.after_speaker_embedding(x)
Anthony Larcher's avatar
Anthony Larcher committed
537

Anthony Larcher's avatar
Anthony Larcher committed
538
539
540
541
542
543
        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
544
        return x
Anthony Larcher's avatar
Anthony Larcher committed
545

Anthony Larcher's avatar
Anthony Larcher committed
546

Anthony Larcher's avatar
Anthony Larcher committed
547
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
548
           dataset_yaml,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
549
           epochs=100,
Anthony Larcher's avatar
Anthony Larcher committed
550
           lr=0.01,
Anthony Larcher's avatar
Anthony Larcher committed
551
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
552
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
553
554
555
556
           loss="cce",
           aam_margin=0.5,
           aam_s=30,
           patience=10,
Anthony Larcher's avatar
Anthony Larcher committed
557
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
558
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
559
           multi_gpu=True,
560
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
561
           opt='sgd',
Anthony Larcher's avatar
Anthony Larcher committed
562
563
           reset_parts=[],
           freeze_parts=[],
Anthony Larcher's avatar
Anthony Larcher committed
564
           num_thread=1):
565
566
    """

Anthony Larcher's avatar
Anthony Larcher committed
567
568
569
570
571
572
    :param speaker_number:
    :param dataset_yaml:
    :param epochs:
    :param lr:
    :param model_yaml:
    :param model_name:
Anthony Larcher's avatar
Anthony Larcher committed
573
574
575
576
    :param loss:
    :param aam_margin:
    :param aam_s:
    :param patience:
Anthony Larcher's avatar
Anthony Larcher committed
577
578
579
580
    :param tmp_model_name:
    :param best_model_name:
    :param multi_gpu:
    :param clipping:
Anthony Larcher's avatar
Anthony Larcher committed
581
582
583
    :param opt:
    :param reset_parts:
    :param freeze_parts:
Anthony Larcher's avatar
Anthony Larcher committed
584
    :param num_thread:
585
586
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
587
588
589
590
    # 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
591

Anthony Larcher's avatar
Anthony Larcher committed
592
    logging.critical(f"Start process at {time.strftime('%H:%M:%S', time.localtime())}")
593

Anthony Larcher's avatar
Anthony Larcher committed
594
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
595

Anthony Larcher's avatar
Anthony Larcher committed
596
597
    # Start from scratch
    if model_name is None:
Anthony Larcher's avatar
Anthony Larcher committed
598
        # Initialize a first model
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
599
        if model_yaml == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
600
            model = Xtractor(speaker_number, "xvector")
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
601
        elif model_yaml == "rawnet2":
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
602
            model = Xtractor(speaker_number, "rawnet2")
Anthony Larcher's avatar
Anthony Larcher committed
603
        else:
Anthony Larcher's avatar
Anthony Larcher committed
604
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
    # 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
620
        new_model_dict = model.state_dict()
Anthony Larcher's avatar
Anthony Larcher committed
621
622
623
624
625
626
627
        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
628

Anthony Larcher's avatar
Anthony Larcher committed
629

Anthony Larcher's avatar
Anthony Larcher committed
630
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
631
632
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
633
634
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
635
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
636
637

    """
Anthony Larcher's avatar
Anthony Larcher committed
638
639
640
641
    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
642
    """
Anthony Larcher's avatar
Anthony Larcher committed
643
644
645
646
    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"])
647

Anthony Larcher's avatar
Anthony Larcher committed
648
    torch.manual_seed(dataset_params['seed'])
649
650
651
    training_set = SideSet(dataset_yaml, 
                           set_type="train", 
                           dataset_df=training_df, 
Anthony Larcher's avatar
Anthony Larcher committed
652
653
                           chunk_per_segment=dataset_params['train']['chunk_per_segment'], 
                           overlap=dataset_params['train']['overlap'])
Anthony Larcher's avatar
Anthony Larcher committed
654
655
656
    training_loader = DataLoader(training_set,
                                 batch_size=dataset_params["batch_size"],
                                 shuffle=True,
Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
657
                                 drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
658
                                 pin_memory=True,
Anthony Larcher's avatar
Anthony Larcher committed
659
                                 num_workers=num_thread)
660

Anthony Larcher's avatar
Anthony Larcher committed
661
662
663
    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
664
                                   drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
665
                                   pin_memory=True,
Anthony Larcher's avatar
Anthony Larcher committed
666
                                   num_workers=num_thread)
Anthony Larcher's avatar
Anthony Larcher committed
667

Anthony Larcher's avatar
Anthony Larcher committed
668
669
670
671
672
673
    # Add for TensorBoard
    #dataiter = iter(training_loader)
    #data, labels = dataiter.next()
    #writer.add_graph(model, data)


Anthony Larcher's avatar
Anthony Larcher committed
674
675
676
    """
    Set the training options
    """
Anthony Larcher's avatar
Anthony Larcher committed
677
    if opt == 'adam':
Anthony Larcher's avatar
Anthony Larcher committed
678
        _optimizer = torch.optim.Adam
Anthony Larcher's avatar
Anthony Larcher committed
679
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
680
681
    elif opt == 'rmsprop':
        _optimizer = torch.optim.RMSprop
Anthony Larcher's avatar
Anthony Larcher committed
682
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
683
684
685
    else: # opt == 'sgd'
        _optimizer = torch.optim.SGD
        _options = {'lr': lr, 'momentum': 0.9}
Anthony Larcher's avatar
Anthony Larcher committed
686

Anthony Larcher's avatar
Anthony Larcher committed
687
688
689
690
691
692
693
694
695
696
697
698
699
700
    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
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
    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
731

Anthony Larcher's avatar
Anthony Larcher committed
732
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', verbose=True)
733

Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
734
    best_accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
735
    best_accuracy_epoch = 1
Anthony Larcher's avatar
Anthony Larcher committed
736
    curr_patience = patience
Anthony Larcher's avatar
Anthony Larcher committed
737
    for epoch in range(1, epochs + 1):
738
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
739
740
741
        if curr_patience == 0:
            print(f"Stopping at epoch {epoch} for cause of patience")
            break
Anthony Larcher's avatar
Anthony Larcher committed
742
743
744
745
746
747
        model = train_epoch(model,
                            epoch,
                            training_loader,
                            optimizer,
                            dataset_params["log_interval"],
                            device=device,
Anthony Larcher's avatar
Anthony Larcher committed
748
749
                            clipping=clipping,
                            tb_writer=writer)
750
751

        # Add the cross validation here
Anthony Larcher's avatar
Anthony Larcher committed
752
        accuracy, val_loss = cross_validation(model, validation_loader, device=device)
Anthony Larcher's avatar
Anthony Larcher committed
753
        logging.critical(f"***{time.strftime('%H:%M:%S', time.localtime())} Cross validation accuracy = {accuracy} %")
754
755
756
757

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

Anthony Larcher's avatar
Anthony Larcher committed
758
759
760
761
        # remember best accuracy and save checkpoint
        is_best = accuracy > best_accuracy
        best_accuracy = max(accuracy, best_accuracy)

Anthony Larcher's avatar
Anthony Larcher committed
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
        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
778
779
780

        if is_best:
            best_accuracy_epoch = epoch
Anthony Larcher's avatar
Anthony Larcher committed
781
782
783
            curr_patience = patience
        else:
            curr_patience -= 1
Anthony Larcher's avatar
Anthony Larcher committed
784
    #writer.close()
785

786
787
788
    for ii in range(torch.cuda.device_count()):
        print(torch.cuda.memory_summary(ii))

Anthony Larcher's avatar
Anthony Larcher committed
789
    logging.critical(f"Best accuracy {best_accuracy * 100.} obtained at epoch {best_accuracy_epoch}")
790

Anthony Larcher's avatar
Anthony Larcher committed
791
def train_epoch(model, epoch, training_loader, optimizer, log_interval, device, clipping=False, tb_writer=None):
792
793
794
795
    """

    :param model:
    :param epoch:
Anthony Larcher's avatar
Anthony Larcher committed
796
    :param training_loader:
797
    :param optimizer:
Anthony Larcher's avatar
Anthony Larcher committed
798
799
800
    :param log_interval:
    :param device:
    :param clipping:
801
802
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
803
    model.train()
Anthony Larcher's avatar
Anthony Larcher committed
804
    criterion = torch.nn.CrossEntropyLoss(reduction='mean')
805
806

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
807
    running_loss = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
808
    for batch_idx, (data, target) in enumerate(training_loader):
809
810
811
        target = target.squeeze()
        optimizer.zero_grad()
        output = model(data.to(device))
Anthony Larcher's avatar
Anthony Larcher committed
812
        #with GuruMeditation():
813
        loss = criterion(output, target.to(device))
Anthony Larcher's avatar
Anthony Larcher committed
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
        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)

843
844
845
846
847
848
849
850
851
852
853
854
855
        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
856
857
858
    return model


Anthony Larcher's avatar
Anthony Larcher committed
859
def cross_validation(model, validation_loader, device):
860
861
862
    """

    :param model:
Anthony Larcher's avatar
Anthony Larcher committed
863
864
    :param validation_loader:
    :param device:
865
866
867
868
869
    :return:
    """
    model.eval()

    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
870
    loss = 0.0
871
    criterion = torch.nn.CrossEntropyLoss()
Anthony Larcher's avatar
Anthony Larcher committed
872
873
874
875
876
877
    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()
878

Anthony Larcher's avatar
Anthony Larcher committed
879
880
            loss += criterion(output, target.to(device))
    
Anthony Larcher's avatar
Anthony Larcher committed
881
882
883
884
    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
885
886
887
888
889
890
891
def extract_embeddings(idmap_name,
                       speaker_number,
                       model_filename,
                       model_yaml,
                       data_root_name ,
                       device,
                       file_extension="wav",
892
893
                       transform_pipeline=None,
                       num_thread=1):
Anthony Larcher's avatar
Anthony Larcher committed
894

Anthony Larcher's avatar
Anthony Larcher committed
895
    if isinstance(idmap_name, IdMap):
896
897
898
899
        idmap = idmap_name
    else:
        idmap = IdMap(idmap_name)

Anthony Larcher's avatar
Anthony Larcher committed
900
    # Create dataset to load the data
Anthony Larcher's avatar
Anthony Larcher committed
901
902
903
904
    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
905

906

907
908
909
910
911
912
    dataloader = DataLoader(dataset,
                            batch_size=1,
                            shuffle=False,
                            drop_last=False,
                            pin_memory=True,
                            num_workers=num_thread)
913

Anthony Larcher's avatar
Anthony Larcher committed
914
    with torch.no_grad():
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
        # Load the model
        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

        model.eval()
        model.to(device)

        # 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]
    
        # Create the StatServer
        embeddings = StatServer()
        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))

        # Process the data
        with torch.no_grad():
            #for idx in tqdm.tqdm(range(len(dataset))):
            for idx, (data, mod, seg, start, stop) in tqdm.tqdm(enumerate(dataloader)):
                #data, mod, seg, start, stop = dataset[idx]
                vec = model(data.to(device), is_eval=True)
                #vec = model(data[None, :, :].to(device), is_eval=True)
                #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
952
953
954
955

    return embeddings


Anthony Larcher's avatar
Anthony Larcher committed
956
957
958
959
960
961
962
963
964
965
966
967
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):

968
969
970
971
972
973
974
975
976
977
978

    # 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
979

980
981
982
        # 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
983
                                    int(sample_rate * (window_length - overlap)))
984
985

        # Create a numpy array to store the current x-vectors
Anthony Larcher's avatar
Anthony Larcher committed
986
        model_names.append(numpy.array([mod + f"_{ii}" for ii in range(len(chunk_starts))]).astype("U"))
987
988
989
990
991
992
993
        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
994
995
996
997
    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)
998
    assert sliding_idmap.validate()
Anthony Larcher's avatar
Anthony Larcher committed
999

Anthony Larcher's avatar
Anthony Larcher committed
1000
    embeddings = extract_embeddings(sliding_idmap,