xvector.py 71.5 KB
Newer Older
Anthony Larcher's avatar
Anthony Larcher committed
1
# coding: utf-8 -*-
Anthony Larcher's avatar
Anthony Larcher committed
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
#
# This file is part of SIDEKIT.
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is free software: you can redistribute it and/or modify
# it under the terms of the GNU LLesser General Public License as
# published by the Free Software Foundation, either version 3 of the License,
# or (at your option) any later version.
#
# SIDEKIT is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU Lesser General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with SIDEKIT.  If not, see <http://www.gnu.org/licenses/>.

"""
Anthony Larcher's avatar
v1.3.7    
Anthony Larcher committed
25
Copyright 2014-2021 Anthony Larcher, Yevhenii Prokopalo
Anthony Larcher's avatar
Anthony Larcher committed
26
"""
Anthony Larcher's avatar
Anthony Larcher committed
27

Anthony Larcher's avatar
Anthony Larcher committed
28

Anthony Larcher's avatar
Anthony Larcher committed
29
import logging
Gaël Le Lan's avatar
Gaël Le Lan committed
30
import math
Anthony Larcher's avatar
Anthony Larcher committed
31
import os
Gaël Le Lan's avatar
Gaël Le Lan committed
32
import pdb
Anthony Larcher's avatar
minor    
Anthony Larcher committed
33
import pickle
Anthony Larcher's avatar
Anthony Larcher committed
34
import shutil
Anthony Larcher's avatar
Anthony Larcher committed
35
import sys
36
import time
Gaël Le Lan's avatar
Gaël Le Lan committed
37
38
39
40
41
42
import traceback
from collections import OrderedDict

import matplotlib.pyplot as plt
import numpy
import pandas
Anthony Larcher's avatar
Anthony Larcher committed
43
import torch
Anthony Larcher's avatar
Anthony Larcher committed
44
import tqdm
Anthony Larcher's avatar
Anthony Larcher committed
45
import yaml
Gaël Le Lan's avatar
Gaël Le Lan committed
46
from ..iv_scoring import cosine_scoring
Gaël Le Lan's avatar
Gaël Le Lan committed
47
48
from sklearn.model_selection import train_test_split
from torch.utils.data import DataLoader
Anthony Larcher's avatar
Anthony Larcher committed
49
from torchvision import transforms
Anthony Larcher's avatar
Anthony Larcher committed
50
from collections import OrderedDict
Gaël Le Lan's avatar
Gaël Le Lan committed
51
from .xsets import SideSet, SpkSet
Anthony Larcher's avatar
debug    
Anthony Larcher committed
52
from .xsets import FileSet
Anthony Larcher's avatar
Anthony Larcher committed
53
from .xsets import IdMapSet
Anthony Larcher's avatar
Anthony Larcher committed
54
from .xsets import IdMapSet_per_speaker
Anthony Larcher's avatar
debug    
Anthony Larcher committed
55
from .xsets import SpkSet
Gaël Le Lan's avatar
Gaël Le Lan committed
56
from .res_net import RawPreprocessor, ResBlockWFMS, ResBlock, PreResNet34, PreFastResNet34
Anthony Larcher's avatar
Anthony Larcher committed
57
from ..bosaris import IdMap
Anthony Larcher's avatar
Anthony Larcher committed
58
59
60
61
from ..bosaris import Key
from ..bosaris import Ndx
from ..bosaris.detplot import rocch
from ..bosaris.detplot import rocch2eer
Anthony Larcher's avatar
Anthony Larcher committed
62
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
63
from ..iv_scoring import cosine_scoring
Anthony Larcher's avatar
Anthony Larcher committed
64
from torch.utils.data import DataLoader
Anthony Larcher's avatar
Anthony Larcher committed
65
from sklearn.model_selection import train_test_split
Anthony Larcher's avatar
Anthony Larcher committed
66
from .sincnet import SincNet
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
67
68
69
from .loss import ArcLinear
from .loss import ArcFace
from .loss import l2_norm
Anthony Larcher's avatar
Anthony Larcher committed
70
from .loss import ArcMarginProduct
Gaël Le Lan's avatar
Gaël Le Lan committed
71
from torch.cuda.amp import autocast, GradScaler
Gaël Le Lan's avatar
Gaël Le Lan committed
72
import torchaudio
Anthony Larcher's avatar
Anthony Larcher committed
73

Anthony Larcher's avatar
ddp    
Anthony Larcher committed
74

Anthony Larcher's avatar
Anthony Larcher committed
75
os.environ['MKL_THREADING_LAYER'] = 'GNU'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
76

Anthony Larcher's avatar
Anthony Larcher committed
77
78
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
v1.3.7    
Anthony Larcher committed
79
__copyright__ = "Copyright 2015-2021 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
80
81
82
83
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
84
85


Anthony Larcher's avatar
Anthony Larcher committed
86
87
logging.basicConfig(format='%(asctime)s %(message)s')

Anthony Larcher's avatar
Anthony Larcher committed
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108

# 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)
Gaël Le Lan's avatar
Gaël Le Lan committed
109
            self.halt(str(value))
Anthony Larcher's avatar
Anthony Larcher committed
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133

    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:
Gaël Le Lan's avatar
Gaël Le Lan committed
134
        plt.imshow(numpy.transpose(npimg, (1, 2, 0)))
Anthony Larcher's avatar
Anthony Larcher committed
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149

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):
    '''
Anthony Larcher's avatar
Anthony Larcher committed
150

Anthony Larcher's avatar
Anthony Larcher committed
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
    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


Gaël Le Lan's avatar
Gaël Le Lan committed
171
172
def test_metrics(model,
                 device,
Gaël Le Lan's avatar
Gaël Le Lan committed
173
174
175
                 speaker_number,
                 num_thread,
                 mixed_precision):
Gaël Le Lan's avatar
Gaël Le Lan committed
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
    """Compute model metrics

    Args:
        model ([type]): [description]
        validation_loader ([type]): [description]
        device ([type]): [description]
        speaker_number ([type]): [description]
        model_archi ([type]): [description]

    Raises:
        NotImplementedError: [description]
        NotImplementedError: [description]

    Returns:
        [type]: [description]
    """
Gaël Le Lan's avatar
Gaël Le Lan committed
192
193
194
    idmap_test_filename = 'h5f/idmap_test.h5'
    ndx_test_filename = 'h5f/ndx_test.h5'
    key_test_filename = 'h5f/key_test.h5'
Gaël Le Lan's avatar
Gaël Le Lan committed
195
    data_root_name='/home/rsgb7088/data/vox1/test/wav'
Gaël Le Lan's avatar
Gaël Le Lan committed
196

Gaël Le Lan's avatar
Gaël Le Lan committed
197
    transform_pipeline = dict()
Gaël Le Lan's avatar
Gaël Le Lan committed
198
199
200
201
202
203
204
205
206
207
208
    #mfcc_config = dict()
    #mfcc_config['nb_filters'] = 81
    #mfcc_config['nb_ceps'] = 80
    #mfcc_config['lowfreq'] = 133.333
    #mfcc_config['maxfreq'] = 6855.4976
    #mfcc_config['win_time'] = 0.025
    #mfcc_config['shift'] = 0.01
    #mfcc_config['n_fft'] = 2048
    #transform_pipeline['MFCC'] = mfcc_config
    #transform_pipeline['CMVN'] = {}
    transform_pipeline = None
Gaël Le Lan's avatar
Gaël Le Lan committed
209
210

    xv_stat = extract_embeddings(idmap_name=idmap_test_filename,
Gaël Le Lan's avatar
Gaël Le Lan committed
211
212
                                 speaker_number=speaker_number,
                                 model_filename=model,
Gaël Le Lan's avatar
Gaël Le Lan committed
213
                                 data_root_name=data_root_name,
Gaël Le Lan's avatar
Gaël Le Lan committed
214
                                 device=device,
Gaël Le Lan's avatar
Gaël Le Lan committed
215
216
217
                                 transform_pipeline=transform_pipeline,
                                 num_thread=num_thread,
                                 mixed_precision=mixed_precision)
Gaël Le Lan's avatar
Gaël Le Lan committed
218

Gaël Le Lan's avatar
Gaël Le Lan committed
219
    scores = cosine_scoring(xv_stat,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
220
221
                            xv_stat,
                            Ndx(ndx_test_filename),
Gaël Le Lan's avatar
Gaël Le Lan committed
222
223
                            wccn=None,
                            check_missing=True)
Gaël Le Lan's avatar
Gaël Le Lan committed
224

Gaël Le Lan's avatar
Gaël Le Lan committed
225
    tar, non = scores.get_tar_non(Key(key_test_filename))
Anthony Larcher's avatar
debug    
Anthony Larcher committed
226

Gaël Le Lan's avatar
Gaël Le Lan committed
227
    test_eer = eer(numpy.array(non).astype(numpy.double), numpy.array(tar).astype(numpy.double))
Gaël Le Lan's avatar
Gaël Le Lan committed
228

Gaël Le Lan's avatar
Gaël Le Lan committed
229
    return test_eer
Gaël Le Lan's avatar
Gaël Le Lan committed
230

Anthony Larcher's avatar
Anthony Larcher committed
231

232
def get_lr(optimizer):
Anthony Larcher's avatar
Anthony Larcher committed
233
234
235
236
237
    """

    :param optimizer:
    :return:
    """
238
239
240
241
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
242
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
243
244
245
246
247
248
249
250
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
251
252
253
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
254

Anthony Larcher's avatar
Anthony Larcher committed
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
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
275

Anthony Larcher's avatar
Anthony Larcher committed
276

Anthony Larcher's avatar
Anthony Larcher committed
277
278
279
280
281
282
283
284
285
286
287
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
288
        super(GruPooling, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
        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
311

Anthony Larcher's avatar
Anthony Larcher committed
312
class Xtractor(torch.nn.Module):
313
314
315
    """
    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
316

Anthony Larcher's avatar
Anthony Larcher committed
317
318
319
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
320
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
321
322
323
                 norm_embedding=False,
                 aam_margin=0.5,
                 aam_s=0.5):
Anthony Larcher's avatar
Anthony Larcher committed
324
325
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
326
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
327
        """
Anthony Larcher's avatar
Anthony Larcher committed
328
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
329
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
330
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
331
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
332

Anthony Larcher's avatar
Anthony Larcher committed
333
334
        print(f"Speaker number : {self.speaker_number}")

Anthony Larcher's avatar
Anthony Larcher committed
335
        if model_archi == "xvector":
336
337
338
339
340
341
342

            self.input_nbdim = 2

            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
343

Gaël Le Lan's avatar
Gaël Le Lan committed
344
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
345
346
347
            self.activation = torch.nn.LeakyReLU(0.2)

            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
348

Anthony Larcher's avatar
xv    
Anthony Larcher committed
349
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
350
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
351
                ("activation1", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
352
                ("batch_norm1", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
353
354
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
355
                ("batch_norm2", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
356
357
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
358
                ("batch_norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
359
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
360
                ("activation4", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
361
                ("batch_norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
362
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
363
                ("activation5", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
364
                ("batch_norm5", torch.nn.BatchNorm1d(1536))
Anthony Larcher's avatar
Anthony Larcher committed
365
366
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
367
            self.stat_pooling = MeanStdPooling()
Gaël Le Lan's avatar
Gaël Le Lan committed
368
            self.stat_pooling_weight_decay = 0
Anthony Larcher's avatar
xv    
Anthony Larcher committed
369
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Gaël Le Lan's avatar
Gaël Le Lan committed
370
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
371
372
            ]))

373
374
            self.embedding_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
375
            if self.loss == "aam":
Gaël Le Lan's avatar
Gaël Le Lan committed
376
                self.after_speaker_embedding = ArcMarginProduct(512,
Anthony Larcher's avatar
Anthony Larcher committed
377
378
379
380
                                                                int(self.speaker_number),
                                                                s=64,
                                                                m=0.2,
                                                                easy_margin=True)
Gaël Le Lan's avatar
Gaël Le Lan committed
381
            elif self.loss == "cce":
Gaël Le Lan's avatar
Gaël Le Lan committed
382
                self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
Gaël Le Lan's avatar
Gaël Le Lan committed
383
                    ("activation6", torch.nn.LeakyReLU(0.2)),
Gaël Le Lan's avatar
Gaël Le Lan committed
384
                    ("batch_norm6", torch.nn.BatchNorm1d(512)),
Gaël Le Lan's avatar
Gaël Le Lan committed
385
                    ("dropout6", torch.nn.Dropout(p=0.05)),
Gaël Le Lan's avatar
Gaël Le Lan committed
386
                    ("linear7", torch.nn.Linear(512, 512)),
Gaël Le Lan's avatar
Gaël Le Lan committed
387
                    ("activation7", torch.nn.LeakyReLU(0.2)),
Gaël Le Lan's avatar
Gaël Le Lan committed
388
389
                    ("batch_norm7", torch.nn.BatchNorm1d(512)),
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
Gaël Le Lan's avatar
Gaël Le Lan committed
390
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
391

392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
            self.sequence_network_weight_decay = 0.0002
            self.before_speaker_embedding_weight_decay = 0.002
            self.after_speaker_embedding_weight_decay = 0.002
            self.embedding_size = 512

        elif model_archi == "resnet34":
            self.input_nbdim = 2
            self.preprocessor = None
            self.sequence_network = PreResNet34()

            self.before_speaker_embedding = torch.nn.Linear(in_features = 5120,
                                                            out_features = 256)

            self.stat_pooling = MeanStdPooling()
            self.stat_pooling_weight_decay = 0

            self.embedding_size = 256

            self.loss = "aam"
            self.after_speaker_embedding = ArcMarginProduct(256,
                                                            int(self.speaker_number),
                                                            s = 30.0,
                                                            m = 0.20,
                                                            easy_margin = True)

            self.preprocessor_weight_decay = 0.000
Gaël Le Lan's avatar
Gaël Le Lan committed
418
            self.sequence_network_weight_decay = 0.000
419
            self.stat_pooling_weight_decay = 0.000
Gaël Le Lan's avatar
Gaël Le Lan committed
420
421
            self.before_speaker_embedding_weight_decay = 0.00
            self.after_speaker_embedding_weight_decay = 0.00
Anthony Larcher's avatar
Anthony Larcher committed
422

423
424
        elif model_archi == "fastresnet34":
            self.input_nbdim = 2
Anthony Larcher's avatar
Anthony Larcher committed
425
            self.preprocessor = None
426

Gaël Le Lan's avatar
Gaël Le Lan committed
427
428
429
430
431
432
433
434
435
436
437
438
439
440
            melkwargs = dict()
            melkwargs['sample_rate'] = 16000
            melkwargs['n_fft'] = 1024 #2048
            melkwargs['f_min'] = 90 #133.333
            melkwargs['f_max'] = 7600 #6855.4976
            melkwargs['win_length'] = 1024 #400
            melkwargs['hop_length'] = 256 #160
            melkwargs['window_fn'] = torch.hann_window
            #melkwargs['power'] = 2
            melkwargs['n_mels'] = 80

            self.sequence_network = PreFastResNet34(melkwargs=melkwargs)

            self.embedding_size = 256
Gaël Le Lan's avatar
Gaël Le Lan committed
441

Gaël Le Lan's avatar
Gaël Le Lan committed
442
            self.before_speaker_embedding = torch.nn.Linear(in_features = 2560,
Gaël Le Lan's avatar
Gaël Le Lan committed
443
                                                            out_features = self.embedding_size)
Anthony Larcher's avatar
Anthony Larcher committed
444
445
446
447
448

            self.stat_pooling = MeanStdPooling()
            self.stat_pooling_weight_decay = 0


449
            self.loss = "aam"
450
            self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
Anthony Larcher's avatar
Anthony Larcher committed
451
                                                            int(self.speaker_number),
Gaël Le Lan's avatar
Gaël Le Lan committed
452
                                                            s = 30,
Gaël Le Lan's avatar
Gaël Le Lan committed
453
                                                            m = 0.2,
Gaël Le Lan's avatar
Gaël Le Lan committed
454
                                                            easy_margin = False)
Anthony Larcher's avatar
Anthony Larcher committed
455
456
457
458
459
460
461

            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
462
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
463
464
465
466
467
468

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

469
470
            self.input_nbdim = 2

Anthony Larcher's avatar
Anthony Larcher committed
471
472
473
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
474
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
475
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
476
477
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494

            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
495
496
497
498
499
            if self.loss == "aam":
                if loss == 'aam':
                    self.after_speaker_embedding = ArcLinear(1024,
                                                             int(self.speaker_number),
                                                             margin=aam_margin, s=aam_s)
Anthony Larcher's avatar
Anthony Larcher committed
500
            elif self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
501
502
503
                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
504

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
505
506
507
508
509
510
            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
511
        else:
Anthony Larcher's avatar
Anthony Larcher committed
512
513
            is_first_resblock = True

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
514
515
516
517
518
519
            if isinstance(model_archi, dict):
                cfg = model_archi
            else:
                # Load Yaml configuration
                with open(model_archi, 'r') as fh:
                    cfg = yaml.load(fh, Loader=yaml.FullLoader)
Anthony Larcher's avatar
Anthony Larcher committed
520

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
521
            self.loss = cfg["training"]["loss"]
Anthony Larcher's avatar
Anthony Larcher committed
522
            if self.loss == "aam":
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
523
524
                self.aam_margin = cfg["training"]["aam_margin"]
                self.aam_s = cfg["training"]["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
525

Anthony Larcher's avatar
Anthony Larcher committed
526
527
528
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
529
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
530
531
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
532
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
533
534
535
536
537
538
539
540
541
542
543
544
                        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
545
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
546
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
547
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
548
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
549
550
551
552
553
554
                                                        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
555
                self.preprocessor_weight_decay = 0.000
Anthony Larcher's avatar
Anthony Larcher committed
556
557

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
558
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
559
            """
Anthony Larcher's avatar
Anthony Larcher committed
560
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
561
562
563
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
            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
579
580
581
582
583
                if k.startswith("lin"):
                    segmental_layers.append((k, torch.nn.Linear(input_size,
                                                                cfg["segmental"][k]["output"])))
                    input_size = cfg["segmental"][k]["output"]

Anthony Larcher's avatar
Anthony Larcher committed
584
585
586
587
588
                elif k.startswith("conv2D"):
                    segmental_layers.append((k, torch.nn.Conv2d(in_channels=1,
                                                                out_channels=entry_conv_out_channels,
                                                                kernel_size=entry_conv_kernel_size,
                                                                padding=3,
Anthony Larcher's avatar
Anthony Larcher committed
589
                                                                stride=1)))
Anthony Larcher's avatar
Anthony Larcher committed
590

Anthony Larcher's avatar
Anthony Larcher committed
591
                elif k.startswith("conv"):
Anthony Larcher's avatar
Anthony Larcher committed
592
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
593
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
594
595
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
596
597
                    input_size = cfg["segmental"][k]["output_channels"]

Anthony Larcher's avatar
Anthony Larcher committed
598
599
                elif k.startswith("ctrans"):
                    segmental_layers.append((k, torch.nn.ConvTranspose1d(input_size,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
600
                                                                         cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
601
602
                                                                         kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                         dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
603
604
605
                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
606
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
Anthony Larcher committed
607
608
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
609
610
611
612
613
614
                elif k.startswith('resblock'):
                    segmental_layers.append((ResBlock(cfg["segmental"][k]["input_channel"],
                                                      cfg["segmental"][k]["output_channel"],
                                                      is_first_resblock)))
                    is_first_resblock = False

Anthony Larcher's avatar
Anthony Larcher committed
615
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
616
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
617

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
618
619
620
621
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
622
            tmp_input_size = input_size * 2
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
623
624
625
626
            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
debug    
Anthony Larcher committed
627
                tmp_input_size = cfg["stat_pooling"]["gru_node"]
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
628

Anthony Larcher's avatar
Anthony Larcher committed
629
630
            self.stat_pooling_weight_decay = cfg["stat_pooling"]["weight_decay"]

Anthony Larcher's avatar
Anthony Larcher committed
631
            """
Anthony Larcher's avatar
Anthony Larcher committed
632
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
633
            """
Anthony Larcher's avatar
Anthony Larcher committed
634
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
635
            input_size = tmp_input_size
Anthony Larcher's avatar
xv    
Anthony Larcher committed
636
637
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
638
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
639
640
                    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
641
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
642
643
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
644
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
645
646

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
647
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
648

Anthony Larcher's avatar
Anthony Larcher committed
649
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
650
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
651
652

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
653
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
654

Anthony Larcher's avatar
Anthony Larcher committed
655
            self.embedding_size = input_size
Anthony Larcher's avatar
Anthony Larcher committed
656
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
657
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
658

Anthony Larcher's avatar
Anthony Larcher committed
659
            # if loss_criteria is "cce"
Anthony Larcher's avatar
xv    
Anthony Larcher committed
660
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
661
662
663
664
665
666
667
668
            if self.loss == "cce":
                after_embedding_layers = []
                for k in cfg["after_embedding"].keys():
                    if k.startswith("lin"):
                        if cfg["after_embedding"][k]["output"] == "speaker_number":
                            after_embedding_layers.append((k, torch.nn.Linear(input_size, self.speaker_number)))
                        else:
                            after_embedding_layers.append((k, torch.nn.Linear(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
669
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
670
                            input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
671

Anthony Larcher's avatar
Anthony Larcher committed
672
673
674
675
676
                    elif k.startswith('arc'):
                        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
677

Anthony Larcher's avatar
Anthony Larcher committed
678
679
                    elif k.startswith("activation"):
                        after_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
680

Anthony Larcher's avatar
Anthony Larcher committed
681
682
                    elif k.startswith('batch_norm'):
                        after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
683

Anthony Larcher's avatar
Anthony Larcher committed
684
685
686
687
688
689
                    elif k.startswith('dropout'):
                        after_embedding_layers.append((k, torch.nn.Dropout(p=cfg["after_embedding"][k])))

                self.after_speaker_embedding = torch.nn.Sequential(OrderedDict(after_embedding_layers))

            elif self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
690
                self.norm_embedding = True
Anthony Larcher's avatar
Anthony Larcher committed
691
692
693
694
695
696
                self.after_speaker_embedding = ArcMarginProduct(input_size,
                                                                int(self.speaker_number),
                                                                s=64,
                                                                m=0.2,
                                                                easy_margin=True)

Anthony Larcher's avatar
arcface    
Anthony Larcher committed
697
698
699
700
                #self.after_speaker_embedding = ArcLinear(input_size,
                #                                         self.speaker_number,
                #                                         margin=self.aam_margin,
                #                                         s=self.aam_s)
Anthony Larcher's avatar
Anthony Larcher committed
701
702
703
704
                #self.after_speaker_embedding = ArcFace(embedding_size=input_size,
                #                                       classnum=self.speaker_number,
                #                                       s=64.,
                #                                       m=0.5)
Anthony Larcher's avatar
Anthony Larcher committed
705

Anthony Larcher's avatar
Anthony Larcher committed
706
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
707

Anthony Larcher's avatar
Anthony Larcher committed
708

709
    def forward(self, x, is_eval=False, target=None, extract_after_pooling=False):
710
711
712
        """

        :param x:
713
        :param is_eval: False for training
714
715
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
716
717
718
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
719
        x = self.sequence_network(x)
720

Anthony Larcher's avatar
Anthony Larcher committed
721
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
722
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
723

724
725
        if extract_after_pooling:
            return x
726

Anthony Larcher's avatar
Anthony Larcher committed
727
        x = self.before_speaker_embedding(x)
728

Anthony Larcher's avatar
Anthony Larcher committed
729
        if self.norm_embedding:
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
730
            #x_norm = x.norm(p=2,dim=1, keepdim=True) / 10. # Why  10. ?
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
731
732
733
            #x_norm = torch.linalg.norm(x, ord=2, dim=1, keepdim=True, out=None, dtype=None)
            #x = torch.div(x, x_norm)
            x = l2_norm(x)
Gaël Le Lan's avatar
Gaël Le Lan committed
734

Anthony Larcher's avatar
Anthony Larcher committed
735
        if self.loss == "cce":
Gaël Le Lan's avatar
Gaël Le Lan committed
736
            if is_eval:
Gaël Le Lan's avatar
Gaël Le Lan committed
737
                return x
Gaël Le Lan's avatar
Gaël Le Lan committed
738
            else:
Gaël Le Lan's avatar
Gaël Le Lan committed
739
                return self.after_speaker_embedding(x), x
Anthony Larcher's avatar
Anthony Larcher committed
740

Anthony Larcher's avatar
Anthony Larcher committed
741
        elif self.loss == "aam":
Gaël Le Lan's avatar
Gaël Le Lan committed
742
743
            if is_eval:
                x = torch.nn.functional.normalize(x, dim=1)
Anthony Larcher's avatar
Anthony Larcher committed
744
            else:
Gaël Le Lan's avatar
Gaël Le Lan committed
745
                x = self.after_speaker_embedding(torch.nn.functional.normalize(x, dim=1), target=target), torch.nn.functional.normalize(x, dim=1)
Gaël Le Lan's avatar
Gaël Le Lan committed
746

Anthony Larcher's avatar
Anthony Larcher committed
747
        return x
Anthony Larcher's avatar
Anthony Larcher committed
748

749
750
    def context_size(self):
        context = 1
Anthony Larcher's avatar
Anthony Larcher committed
751
752
753
754
755
756
757
758
759
        if isinstance(self, Xtractor):
            for name, module in self.sequence_network.named_modules():
                if name.startswith("conv"):
                    context += module.dilation[0] * (module.kernel_size[0] - 1)
        else:
            for name, module in self.module.sequence_network.named_modules():
                if name.startswith("conv"):
                    context += module.dilation[0] * (module.kernel_size[0] - 1)
        return context
Anthony Larcher's avatar
Anthony Larcher committed
760

Anthony Larcher's avatar
Anthony Larcher committed
761
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
762
           dataset_yaml,
Anthony Larcher's avatar
Anthony Larcher committed
763
764
           epochs=None,
           lr=None,
Anthony Larcher's avatar
Anthony Larcher committed
765
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
766
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
767
768
769
770
           loss=None,
           aam_margin=None,
           aam_s=None,
           patience=None,
Anthony Larcher's avatar
Anthony Larcher committed
771
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
772
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
773
           multi_gpu=True,
Gaël Le Lan's avatar
Gaël Le Lan committed
774
           mixed_precision=False,
775
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
776
           opt=None,
Anthony Larcher's avatar
Anthony Larcher committed
777
778
           reset_parts=[],
           freeze_parts=[],
Anthony Larcher's avatar
Anthony Larcher committed
779
780
781
782
           num_thread=None,
           write_batches_to_disk=False,
           load_batches_from_disk=False,
           tmp_batch_dir=None):
783
784
    """

Anthony Larcher's avatar
Anthony Larcher committed
785
786
787
788
789
790
    :param speaker_number:
    :param dataset_yaml:
    :param epochs:
    :param lr:
    :param model_yaml:
    :param model_name:
Anthony Larcher's avatar
Anthony Larcher committed
791
792
793
794
    :param loss:
    :param aam_margin:
    :param aam_s:
    :param patience:
Anthony Larcher's avatar
Anthony Larcher committed
795
796
797
798
    :param tmp_model_name:
    :param best_model_name:
    :param multi_gpu:
    :param clipping:
Anthony Larcher's avatar
Anthony Larcher committed
799
800
801
    :param opt:
    :param reset_parts:
    :param freeze_parts:
Anthony Larcher's avatar
Anthony Larcher committed
802
    :param num_thread:
803
804
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
805
806
807
808
    # 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
809

Anthony Larcher's avatar
Anthony Larcher committed
810
811
812
    if write_batches_to_disk:
        load_batches_from_disk = True

Anthony Larcher's avatar
Anthony Larcher committed
813
    if num_thread is None:
Gaël Le Lan's avatar
Gaël Le Lan committed
814
815
        import multiprocessing

Anthony Larcher's avatar
Anthony Larcher committed
816
817
        num_thread = multiprocessing.cpu_count()

Anthony Larcher's avatar
Anthony Larcher committed
818
    logging.critical(f"Use {num_thread} cpus")
Anthony Larcher's avatar
Anthony Larcher committed
819
    logging.critical(f"Start process at {time.strftime('%H:%M:%S', time.localtime())}")
820

Anthony Larcher's avatar
Anthony Larcher committed
821
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
822
    # Start from scratch
Gaël Le Lan's avatar
Gaël Le Lan committed
823
    if model_name is None and model_yaml in ["xvector", "rawnet2", "resnet34", "fastresnet34"]:
Anthony Larcher's avatar
Anthony Larcher committed
824
        # Initialize a first model
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
825
        if model_yaml == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
826
            model = Xtractor(speaker_number, "xvector", loss=loss)
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
827
        elif model_yaml == "rawnet2":
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
828
            model = Xtractor(speaker_number, "rawnet2")
829
830
        elif model_yaml == "resnet34":
            model = Xtractor(speaker_number, "resnet34")
Gaël Le Lan's avatar
Gaël Le Lan committed
831
832
        elif model_yaml == "fastresnet34":
            model = Xtractor(speaker_number, "fastresnet34")
Gaël Le Lan's avatar
Gaël Le Lan committed
833
        model_archi = model_yaml
Anthony Larcher's avatar
Anthony Larcher committed
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
    else:
        with open(model_yaml, 'r') as fh:
            model_archi = yaml.load(fh, Loader=yaml.FullLoader)
            if epochs is None:
                epochs = model_archi["training"]["epochs"]
            if patience is None:
                patience = model_archi["training"]["patience"]
            if opt is None:
                opt = model_archi["training"]["opt"]
            if lr is None:
                lr = model_archi["training"]["lr"]
            if loss is None:
                loss = model_archi["training"]["loss"]
            if aam_margin is None and model_archi["training"]["loss"] == "aam":
                aam_margin = model_archi["training"]["aam_margin"]
            if aam_s is None and model_archi["training"]["loss"] == "aam":
                aam_s = model_archi["training"]["aam_s"]
            if tmp_model_name is None:
                tmp_model_name = model_archi["training"]["tmp_model_name"]
            if best_model_name is None:
                best_model_name = model_archi["training"]["best_model_name"]
            if multi_gpu is None:
                multi_gpu = model_archi["training"]["multi_gpu"]
            if clipping is None:
                clipping = model_archi["training"]["clipping"]

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
860
        if model_name is None:
Anthony Larcher's avatar
Anthony Larcher committed
861
862
863
            model = Xtractor(speaker_number, model_yaml)

         # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
864
        else:
Anthony Larcher's avatar
Anthony Larcher committed
865
866
867
            # Load the model
            logging.critical(f"*** Load model from = {model_name}")
            checkpoint = torch.load(model_name)
Anthony Larcher's avatar
Anthony Larcher committed
868
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
869

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
870
871
            """
            Here we remove all layers that we don't want to reload
Anthony Larcher's avatar
Anthony Larcher committed
872
        
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
873
874
875
876
877
878
879
880
881
882
883
884
885
            """
            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)}

            new_model_dict = model.state_dict()
            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
886

Gaël Le Lan's avatar
Gaël Le Lan committed
887
888
889
890
    logging.critical(model)

    logging.critical("model_parameters_count: {:d}".format(
        sum(p.numel()
Gaël Le Lan's avatar
Gaël Le Lan committed
891
892
893
894
            for p in model.sequence_network.parameters()
            if p.requires_grad) + \
        sum(p.numel()
            for p in model.before_speaker_embedding.parameters()
Gaël Le Lan's avatar
Gaël Le Lan committed
895
896
            if p.requires_grad)))

Anthony Larcher's avatar
Anthony Larcher committed
897
    embedding_size = model.embedding_size
Anthony Larcher's avatar
Anthony Larcher committed
898

Anthony Larcher's avatar
Anthony Larcher committed
899
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
900
901
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
902
        #model = DDP(model)
Anthony Larcher's avatar
Anthony Larcher committed
903
904
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
905
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
906

Anthony Larcher's avatar
debug    
Anthony Larcher committed
907
908
909
910
    with open(dataset_yaml, "r") as fh:
        dataset_params = yaml.load(fh, Loader=yaml.FullLoader)
        df = pandas.read_csv(dataset_params["dataset_description"])

Anthony Larcher's avatar
Anthony Larcher committed
911
912
913
    if load_batches_from_disk:
        train_batch_fn_format = tmp_batch_dir + "/train/train_{}_batch.h5"
        val_batch_fn_format = tmp_batch_dir + "/val/val_{}_batch.h5"
914

Anthony Larcher's avatar
Anthony Larcher committed
915
916
917
918
919
920
921
    if not load_batches_from_disk or write_batches_to_disk:
        """
        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
922

Anthony Larcher's avatar
minor    
Anthony Larcher committed
923
        if write_batches_to_disk or dataset_params["batch_size"] > 1:
Anthony Larcher's avatar
Anthony Larcher committed
924
925
926
927
            output_format = "numpy"
        else:
            output_format = "pytorch"

Anthony Larcher's avatar
Anthony Larcher committed
928
929
        training_df, validation_df = train_test_split(df, test_size=dataset_params["validation_ratio"])
        torch.manual_seed(dataset_params['seed'])
Gaël Le Lan's avatar
Gaël Le Lan committed
930
        training_set = SpkSet(dataset_yaml,
Gaël Le Lan's avatar
Gaël Le Lan committed
931
932
933
                              set_type="train",
                              dataset_df=training_df,
                              overlap=dataset_params['train']['overlap'],
Gaël Le Lan's avatar
Gaël Le Lan committed
934
935
                              output_format=output_format,
                              windowed=True)
Gaël Le Lan's avatar
Gaël Le Lan committed
936

Gaël Le Lan's avatar
bugfix    
Gaël Le Lan committed
937
        validation_set = SideSet(dataset_yaml,
Gaël Le Lan's avatar
Gaël Le Lan committed
938
939
940
941
                                 set_type="validation",
                                 dataset_df=validation_df,
                                 output_format=output_format)

Anthony Larcher's avatar
Anthony Larcher committed
942

Anthony Larcher's avatar
Anthony Larcher committed
943
        if write_batches_to_disk:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
944
            logging.critical("Start writing batches on disk")
Anthony Larcher's avatar
Anthony Larcher committed
945
946
            training_set.write_to_disk(dataset_params["batch_size"], train_batch_fn_format, num_thread)
            validation_set.write_to_disk(dataset_params["batch_size"], val_batch_fn_format, num_thread)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
947
            logging.critical("---> Done")
Anthony Larcher's avatar
Anthony Larcher committed
948
949

    if load_batches_from_disk:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
950
951
952
953
954
955
956
957
958
959
960
961
962
        training_set = FileSet(train_batch_fn_format)
        validation_set = FileSet(train_batch_fn_format)
        batch_size = 1
    else:
        batch_size = dataset_params["batch_size"]


    print(f"Size of batches = {batch_size}")
    training_loader = DataLoader(training_set,
                                 batch_size=batch_size,
                                 shuffle=True,
                                 drop_last=True,
                                 pin_memory=True,
Gaël Le Lan's avatar
Gaël Le Lan committed
963
964
                                 num_workers=num_thread,
                                 persistent_workers=True)
Anthony Larcher's avatar
debug    
Anthony Larcher committed
965
966
967

    validation_loader = DataLoader(validation_set,
                                   batch_size=batch_size,
Gaël Le Lan's avatar
Gaël Le Lan committed
968
                                   drop_last=False,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
969
                                   pin_memory=True,
Anthony Larcher's avatar
spkset    
Anthony Larcher committed
970
                                   num_workers=num_thread,
Gaël Le Lan's avatar
Gaël Le Lan committed
971
                                   persistent_workers=False)
Anthony Larcher's avatar
Anthony Larcher committed
972

Anthony Larcher's avatar
Anthony Larcher committed
973
974
975
    """
    Set the training options
    """
Anthony Larcher's avatar
Anthony Larcher committed
976
    if opt == 'adam':
Anthony Larcher's avatar
Anthony Larcher committed
977
        _optimizer = torch.optim.Adam
Anthony Larcher's avatar
Anthony Larcher committed
978
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
979
980
    elif opt == 'rmsprop':
        _optimizer = torch.optim.RMSprop
Anthony Larcher's avatar
Anthony Larcher committed
981
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
982
983
984
    else: # opt == 'sgd'
        _optimizer = torch.optim.SGD
        _options = {'lr': lr, 'momentum': 0.9}
Anthony Larcher's avatar
Anthony Larcher committed
985

Anthony Larcher's avatar
Anthony Larcher committed
986
    param_list = []
Anthony Larcher's avatar
Anthony Larcher committed
987
    if type(model) is Xtractor:
Anthony Larcher's avatar
Anthony Larcher committed
988
989
990
991
992
993
994
        if model.preprocessor is not None:
            param_list.append({'params': model.preprocessor.parameters(), 'weight_decay': model.preprocessor_weight_decay})
        param_list.append({'params': model.sequence_network.parameters(), 'weight_decay': model.sequence_network_weight_decay})
        param_list.append({'params': model.stat_pooling.parameters(), 'weight_decay': model.stat_pooling_weight_decay})
        param_list.append({'params': model.before_speaker_embedding.parameters(), 'weight_decay': model.before_speaker_embedding_weight_decay})
        param_list.append({'params': model.after_speaker_embedding.parameters(), 'weight_decay': model.after_speaker_embedding_weight_decay})

Anthony Larcher's avatar
Anthony Larcher committed
995
    else:
Anthony Larcher's avatar
Anthony Larcher committed
996
997
998
999
1000
        if model.module.preprocessor is not None:
            param_list.append({'params': model.module.preprocessor.parameters(), 'weight_decay': model.module.preprocessor_weight_decay})
        param_list.append({'params': model.module.sequence_network.parameters(), 'weight_decay': model.module.sequence_network_weight_decay})
        param_list.append({'params': model.module.stat_pooling.parameters(), 'weight_decay': model.module.stat_pooling_weight_decay})
        param_list.append({'params': model.module.before_speaker_embedding.parameters(), 'weight_decay': model.module.before_speaker_embedding_weight_decay})
For faster browsing, not all history is shown. View entire blame