xvector.py 63.1 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
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 multiprocessing
Anthony Larcher's avatar
Anthony Larcher committed
34
import os
Anthony Larcher's avatar
Anthony Larcher committed
35
import numpy
Anthony Larcher's avatar
Anthony Larcher committed
36
import pandas
Anthony Larcher's avatar
minor    
Anthony Larcher committed
37
import pickle
Anthony Larcher's avatar
Anthony Larcher committed
38
import shutil
Anthony Larcher's avatar
Anthony Larcher committed
39
import sys
40
import time
Anthony Larcher's avatar
Anthony Larcher committed
41
import torch
Anthony Larcher's avatar
Anthony Larcher committed
42
import tqdm
Anthony Larcher's avatar
Anthony Larcher committed
43
44
import yaml

Anthony Larcher's avatar
Anthony Larcher committed
45
from torchvision import transforms
Anthony Larcher's avatar
Anthony Larcher committed
46
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
47
from .xsets import SideSet
Anthony Larcher's avatar
debug    
Anthony Larcher committed
48
from .xsets import FileSet
Anthony Larcher's avatar
Anthony Larcher committed
49
from .xsets import IdMapSet
Anthony Larcher's avatar
Anthony Larcher committed
50
from .xsets import IdMapSet_per_speaker
Anthony Larcher's avatar
Anthony Larcher committed
51
from .res_net import RawPreprocessor, ResBlockWFMS
Anthony Larcher's avatar
Anthony Larcher committed
52
from ..bosaris import IdMap
Anthony Larcher's avatar
Anthony Larcher committed
53
54
55
56
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
57
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
58
from ..iv_scoring import cosine_scoring
Anthony Larcher's avatar
Anthony Larcher committed
59
from torch.utils.data import DataLoader
Anthony Larcher's avatar
Anthony Larcher committed
60
from sklearn.model_selection import train_test_split
Anthony Larcher's avatar
Anthony Larcher committed
61
from .sincnet import SincNet
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
62
63
64
from .loss import ArcLinear
from .loss import ArcFace
from .loss import l2_norm
Anthony Larcher's avatar
Anthony Larcher committed
65
from .loss import ArcMarginProduct
Anthony Larcher's avatar
Anthony Larcher committed
66

Anthony Larcher's avatar
ddp    
Anthony Larcher committed
67
68


Anthony Larcher's avatar
Anthony Larcher committed
69
os.environ['MKL_THREADING_LAYER'] = 'GNU'
Anthony Larcher's avatar
debug    
Anthony Larcher committed
70

Anthony Larcher's avatar
Anthony Larcher committed
71
72
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
v1.3.7    
Anthony Larcher committed
73
__copyright__ = "Copyright 2015-2021 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
74
75
76
77
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
78
79


Anthony Larcher's avatar
Anthony Larcher committed
80
81
logging.basicConfig(format='%(asctime)s %(message)s')

Anthony Larcher's avatar
Anthony Larcher committed
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102

# 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)
Anthony Larcher's avatar
Anthony Larcher committed
103
            self.halt(str(value))
Anthony Larcher's avatar
Anthony Larcher committed
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127

    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:
Anthony Larcher's avatar
Anthony Larcher committed
128
        plt.imshow(numpy.transpose(npimg, (1, 2, 0)))
Anthony Larcher's avatar
Anthony Larcher committed
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143

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
144

Anthony Larcher's avatar
Anthony Larcher committed
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
    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


Anthony Larcher's avatar
Anthony Larcher committed
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
def compute_metrics(model,
                    validation_loader,
                    device,
                    val_embs_shape,
                    speaker_number,
                    model_archi):
    """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]
    """
    val_acc, val_loss, val_eer = cross_validation(model, validation_loader, device, val_embs_shape)
Anthony Larcher's avatar
Anthony Larcher committed
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
    #xv_stat = extract_embeddings(idmap_name='h5f/idmap_test.h5',
    #                             speaker_number=speaker_number,
    #                             model_filename=model,
    #                             model_yaml=model_archi,
    #                             data_root_name="data/vox1/wav/" ,
    #                             device=device,
    #                             transform_pipeline="MFCC,CMVN")

    #scores = cosine_scoring(xv_stat, xv_stat,
    #                        Ndx('h5f/ndx_test.h5'),
    #                        wccn=None, check_missing=True)

    #tar, non = scores.get_tar_non(Key('h5f/key_test.h5'))
    #pmiss, pfa = rocch(numpy.array(tar).astype(numpy.double), numpy.array(non).astype(numpy.double))
    #test_eer = rocch2eer(pmiss, pfa)

    test_eer = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
205
206
    return val_acc, val_loss, val_eer, test_eer

Anthony Larcher's avatar
Anthony Larcher committed
207

208
def get_lr(optimizer):
Anthony Larcher's avatar
Anthony Larcher committed
209
210
211
212
213
    """

    :param optimizer:
    :return:
    """
214
215
216
217
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
218
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
219
220
221
222
223
224
225
226
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
227
228
229
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
230

Anthony Larcher's avatar
Anthony Larcher committed
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
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
251

Anthony Larcher's avatar
Anthony Larcher committed
252

Anthony Larcher's avatar
Anthony Larcher committed
253
254
255
256
257
258
259
260
261
262
263
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
264
        super(GruPooling, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
        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
287

Anthony Larcher's avatar
Anthony Larcher committed
288
class Xtractor(torch.nn.Module):
289
290
291
    """
    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
292

Anthony Larcher's avatar
Anthony Larcher committed
293
294
295
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
296
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
297
298
299
                 norm_embedding=False,
                 aam_margin=0.5,
                 aam_s=0.5):
Anthony Larcher's avatar
Anthony Larcher committed
300
301
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
302
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
303
        """
Anthony Larcher's avatar
Anthony Larcher committed
304
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
305
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
306
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
307
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
308

Anthony Larcher's avatar
Anthony Larcher committed
309
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
310
311
312
313
314
315

            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
316
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
317
318
319
            self.activation = torch.nn.LeakyReLU(0.2)

            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
320

Anthony Larcher's avatar
xv    
Anthony Larcher committed
321
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
322
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
323
                ("activation1", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
324
                ("batch_norm1", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
325
326
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
327
                ("batch_norm2", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
328
329
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
330
                ("batch_norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
331
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
332
                ("activation4", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
333
                ("batch_norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
334
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
335
                ("activation5", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
336
                ("batch_norm5", torch.nn.BatchNorm1d(1536))
Anthony Larcher's avatar
Anthony Larcher committed
337
338
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
339
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
340
            self.stat_pooling_weight_decay = 0
Anthony Larcher's avatar
xv    
Anthony Larcher committed
341
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
342
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
343
344
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
345
            if self.loss == "aam":
Anthony Larcher's avatar
Anthony Larcher committed
346
347
348
349
350
                self.after_speaker_embedding = ArcMarginProduct(512,
                                                                int(self.speaker_number),
                                                                s=64,
                                                                m=0.2,
                                                                easy_margin=True)
Anthony Larcher's avatar
Anthony Larcher committed
351
352
353
            elif self.loss == "cce":
                self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
                    ("activation6", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
354
                    ("batch_norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
355
356
357
                    ("dropout6", torch.nn.Dropout(p=0.05)),
                    ("linear7", torch.nn.Linear(512, 512)),
                    ("activation7", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
358
                    ("batch_norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
359
360
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
361

Anthony Larcher's avatar
Anthony Larcher committed
362
363
364
365
            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
366
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
367
368
369
370
371
372

            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
373
374
375
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
376
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
377
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
378
379
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396

            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
397
398
399
400
401
            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
402
            elif self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
403
404
405
                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
406

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
407
408
409
410
411
412
            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
413
        else:
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
414
415
416
417
418
419
            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
420

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
421
            self.loss = cfg["training"]["loss"]
Anthony Larcher's avatar
Anthony Larcher committed
422
            if self.loss == "aam":
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
423
424
                self.aam_margin = cfg["training"]["aam_margin"]
                self.aam_s = cfg["training"]["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
425

Anthony Larcher's avatar
Anthony Larcher committed
426
427
428
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
429
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
430
431
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
432
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
433
434
435
436
437
438
439
440
441
442
443
444
                        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
445
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
446
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
447
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
448
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
449
450
451
452
453
454
                                                        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
455
                self.preprocessor_weight_decay = 0.000
Anthony Larcher's avatar
Anthony Larcher committed
456
457

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
458
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
459
            """
Anthony Larcher's avatar
Anthony Larcher committed
460
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
461
462
463
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
            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
479
480
481
482
483
484
                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
485
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
486
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
487
488
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
489
490
                    input_size = cfg["segmental"][k]["output_channels"]

Anthony Larcher's avatar
Anthony Larcher committed
491
492
                elif k.startswith("ctrans"):
                    segmental_layers.append((k, torch.nn.ConvTranspose1d(input_size,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
493
                                                                         cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
494
495
                                                                         kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                         dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
496
497
498
                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
499
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
Anthony Larcher committed
500
501
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
502
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
503
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
504

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
505
506
507
508
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
509
            tmp_input_size = input_size * 2
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
510
511
512
513
            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
514
                tmp_input_size = cfg["stat_pooling"]["gru_node"]
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
515

Anthony Larcher's avatar
Anthony Larcher committed
516
517
            self.stat_pooling_weight_decay = cfg["stat_pooling"]["weight_decay"]

Anthony Larcher's avatar
Anthony Larcher committed
518
            """
Anthony Larcher's avatar
Anthony Larcher committed
519
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
520
            """
Anthony Larcher's avatar
Anthony Larcher committed
521
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
522
            input_size = tmp_input_size
Anthony Larcher's avatar
xv    
Anthony Larcher committed
523
524
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
525
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
526
527
                    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
528
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
529
530
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
531
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
532
533

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
534
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
535

Anthony Larcher's avatar
Anthony Larcher committed
536
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
537
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
538
539

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
540
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
541

Anthony Larcher's avatar
Anthony Larcher committed
542
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
543
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
544

Anthony Larcher's avatar
Anthony Larcher committed
545
            # if loss_criteria is "cce"
Anthony Larcher's avatar
xv    
Anthony Larcher committed
546
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
547
548
549
550
551
552
553
554
            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
555
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
556
                            input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
557

Anthony Larcher's avatar
Anthony Larcher committed
558
559
560
561
562
                    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
563

Anthony Larcher's avatar
Anthony Larcher committed
564
565
                    elif k.startswith("activation"):
                        after_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
566

Anthony Larcher's avatar
Anthony Larcher committed
567
568
                    elif k.startswith('batch_norm'):
                        after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
569

Anthony Larcher's avatar
Anthony Larcher committed
570
571
572
573
574
575
                    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
576
577
578
                print(f"input_size = {input_size}")
                self.after_speaker_embedding = ArcMarginProduct(input_size, int(self.speaker_number), s=64, m=0.2, easy_margin=True)
                #self.norm_embedding = True
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
579
580
581
582
                #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
583
584
585
586
                #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
587

Anthony Larcher's avatar
Anthony Larcher committed
588
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
589

Anthony Larcher's avatar
Anthony Larcher committed
590

Anthony Larcher's avatar
Anthony Larcher committed
591
    def forward(self, x, is_eval=False, target=None):
592
593
594
        """

        :param x:
Anthony Larcher's avatar
Anthony Larcher committed
595
        :param is_eval:
596
597
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
598
599
600
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
601
        x = self.sequence_network(x)
602

Anthony Larcher's avatar
Anthony Larcher committed
603
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
604
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
605
606

        x = self.before_speaker_embedding(x)
607

Anthony Larcher's avatar
Anthony Larcher committed
608
        if self.norm_embedding:
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
609
            #x_norm = x.norm(p=2,dim=1, keepdim=True) / 10. # Why  10. ?
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
610
611
612
            #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)
Anthony Larcher's avatar
Anthony Larcher committed
613

Anthony Larcher's avatar
Anthony Larcher committed
614
        if self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
615
616
617
618
            if is_eval:
                return self.after_speaker_embedding(x), x
            else:
                return self.after_speaker_embedding(x)
Anthony Larcher's avatar
Anthony Larcher committed
619

Anthony Larcher's avatar
Anthony Larcher committed
620
621
        elif self.loss == "aam":
            if not is_eval:
Anthony Larcher's avatar
Anthony Larcher committed
622
                x = self.after_speaker_embedding(l2_norm(x), target=target), l2_norm(x)
Anthony Larcher's avatar
Anthony Larcher committed
623
            else:
Anthony Larcher's avatar
Anthony Larcher committed
624
                x = self.after_speaker_embedding(l2_norm(x), target=None), l2_norm(x)
Anthony Larcher's avatar
Anthony Larcher committed
625

Anthony Larcher's avatar
Anthony Larcher committed
626
        return x
Anthony Larcher's avatar
Anthony Larcher committed
627

628
629
    def context_size(self):
        context = 1
Anthony Larcher's avatar
Anthony Larcher committed
630
631
632
633
634
635
636
637
638
        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
639

Anthony Larcher's avatar
Anthony Larcher committed
640
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
641
           dataset_yaml,
Anthony Larcher's avatar
Anthony Larcher committed
642
643
           epochs=None,
           lr=None,
Anthony Larcher's avatar
Anthony Larcher committed
644
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
645
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
646
647
648
649
           loss=None,
           aam_margin=None,
           aam_s=None,
           patience=None,
Anthony Larcher's avatar
Anthony Larcher committed
650
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
651
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
652
           multi_gpu=True,
653
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
654
           opt=None,
Anthony Larcher's avatar
Anthony Larcher committed
655
656
           reset_parts=[],
           freeze_parts=[],
Anthony Larcher's avatar
Anthony Larcher committed
657
658
659
660
           num_thread=None,
           write_batches_to_disk=False,
           load_batches_from_disk=False,
           tmp_batch_dir=None):
661
662
    """

Anthony Larcher's avatar
Anthony Larcher committed
663
664
665
666
667
668
    :param speaker_number:
    :param dataset_yaml:
    :param epochs:
    :param lr:
    :param model_yaml:
    :param model_name:
Anthony Larcher's avatar
Anthony Larcher committed
669
670
671
672
    :param loss:
    :param aam_margin:
    :param aam_s:
    :param patience:
Anthony Larcher's avatar
Anthony Larcher committed
673
674
675
676
    :param tmp_model_name:
    :param best_model_name:
    :param multi_gpu:
    :param clipping:
Anthony Larcher's avatar
Anthony Larcher committed
677
678
679
    :param opt:
    :param reset_parts:
    :param freeze_parts:
Anthony Larcher's avatar
Anthony Larcher committed
680
    :param num_thread:
681
682
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
683
684
685
686
    # 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
687

Anthony Larcher's avatar
Anthony Larcher committed
688
689
690
    if write_batches_to_disk:
        load_batches_from_disk = True

Anthony Larcher's avatar
Anthony Larcher committed
691
692
693
    if num_thread is None:
        num_thread = multiprocessing.cpu_count()

Anthony Larcher's avatar
Anthony Larcher committed
694
    logging.critical(f"Use {num_thread} cpus")
Anthony Larcher's avatar
Anthony Larcher committed
695
    logging.critical(f"Start process at {time.strftime('%H:%M:%S', time.localtime())}")
696

Anthony Larcher's avatar
Anthony Larcher committed
697
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
698
    # Start from scratch
Anthony Larcher's avatar
Anthony Larcher committed
699
    if model_name is None and model_yaml in ["xvector", "rawnet2"]:
Anthony Larcher's avatar
Anthony Larcher committed
700
        # Initialize a first model
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
701
        if model_yaml == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
702
            model = Xtractor(speaker_number, "xvector")
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
703
        elif model_yaml == "rawnet2":
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
704
            model = Xtractor(speaker_number, "rawnet2")
Anthony Larcher's avatar
Anthony Larcher committed
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
    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
731
        if model_name is None:
Anthony Larcher's avatar
Anthony Larcher committed
732
733
734
            model = Xtractor(speaker_number, model_yaml)

         # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
735
        else:
Anthony Larcher's avatar
Anthony Larcher committed
736
737
738
            # Load the model
            logging.critical(f"*** Load model from = {model_name}")
            checkpoint = torch.load(model_name)
Anthony Larcher's avatar
Anthony Larcher committed
739
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
740

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
741
742
            """
            Here we remove all layers that we don't want to reload
Anthony Larcher's avatar
Anthony Larcher committed
743
        
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
744
745
746
747
748
749
750
751
752
753
754
755
756
            """
            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
757

Anthony Larcher's avatar
Anthony Larcher committed
758
    print(model)
Anthony Larcher's avatar
Anthony Larcher committed
759

Anthony Larcher's avatar
Anthony Larcher committed
760
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
761
762
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
763
        #model = DDP(model)
Anthony Larcher's avatar
Anthony Larcher committed
764
765
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
766
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
767

Anthony Larcher's avatar
debug    
Anthony Larcher committed
768
769
770
771
    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
772
773
774
    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"
775

Anthony Larcher's avatar
Anthony Larcher committed
776
777
778
779
780
781
782
    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
783

Anthony Larcher's avatar
minor    
Anthony Larcher committed
784
        if write_batches_to_disk or dataset_params["batch_size"] > 1:
Anthony Larcher's avatar
Anthony Larcher committed
785
786
787
788
            output_format = "numpy"
        else:
            output_format = "pytorch"

Anthony Larcher's avatar
Anthony Larcher committed
789
790
791
792
793
794
        training_df, validation_df = train_test_split(df, test_size=dataset_params["validation_ratio"])
        torch.manual_seed(dataset_params['seed'])
        training_set = SideSet(dataset_yaml,
                               set_type="train",
                               dataset_df=training_df,
                               chunk_per_segment=dataset_params['train']['chunk_per_segment'],
Anthony Larcher's avatar
Anthony Larcher committed
795
796
                               overlap=dataset_params['train']['overlap'],
                               output_format=output_format)
Anthony Larcher's avatar
Anthony Larcher committed
797

Anthony Larcher's avatar
Anthony Larcher committed
798
799
800
801
        validation_set = SideSet(dataset_yaml,
                                 set_type="validation",
                                 dataset_df=validation_df,
                                 output_format=output_format)
Anthony Larcher's avatar
Anthony Larcher committed
802
803


Anthony Larcher's avatar
Anthony Larcher committed
804
        if write_batches_to_disk:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
805
            logging.critical("Start writing batches on disk")
Anthony Larcher's avatar
Anthony Larcher committed
806
807
            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
808
            logging.critical("---> Done")
Anthony Larcher's avatar
Anthony Larcher committed
809
810

    if load_batches_from_disk:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
        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,
                                 num_workers=num_thread)

    validation_loader = DataLoader(validation_set,
                                   batch_size=batch_size,
                                   drop_last=True,
                                   pin_memory=True,
                                   num_workers=num_thread)

Anthony Larcher's avatar
Anthony Larcher committed
832

Anthony Larcher's avatar
Anthony Larcher committed
833
834
835
    """
    Set the training options
    """
Anthony Larcher's avatar
Anthony Larcher committed
836
    if opt == 'adam':
Anthony Larcher's avatar
Anthony Larcher committed
837
        _optimizer = torch.optim.Adam
Anthony Larcher's avatar
Anthony Larcher committed
838
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
839
840
    elif opt == 'rmsprop':
        _optimizer = torch.optim.RMSprop
Anthony Larcher's avatar
Anthony Larcher committed
841
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
842
843
844
    else: # opt == 'sgd'
        _optimizer = torch.optim.SGD
        _options = {'lr': lr, 'momentum': 0.9}
Anthony Larcher's avatar
Anthony Larcher committed
845

Anthony Larcher's avatar
debug    
Anthony Larcher committed
846
847
848
849
850
851
852
853
854
855
856
857
858
    #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
859

Anthony Larcher's avatar
Anthony Larcher committed
860
    param_list = []
Anthony Larcher's avatar
Anthony Larcher committed
861
    if type(model) is Xtractor:
Anthony Larcher's avatar
Anthony Larcher committed
862
863
864
865
866
867
868
        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
869
    else:
Anthony Larcher's avatar
Anthony Larcher committed
870
871
872
873
874
875
876
877
878
        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})
        param_list.append({'params': model.module.after_speaker_embedding.parameters(), 'weight_decay': model.module.after_speaker_embedding_weight_decay})


    optimizer = _optimizer(param_list, **_options)
Anthony Larcher's avatar
Anthony Larcher committed
879
880
881
882
883
884

    #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
885

Anthony Larcher's avatar
Anthony Larcher committed
886
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', verbose=True)
887

Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
888
    best_accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
889
    best_accuracy_epoch = 1
Anthony Larcher's avatar
Anthony Larcher committed
890
    curr_patience = patience
Anthony Larcher's avatar
Anthony Larcher committed
891
892
893
894

    val_acc, val_loss, val_eer, test_eer = compute_metrics(model,
                                                           validation_loader,
                                                           device,
Anthony Larcher's avatar
Anthony Larcher committed
895
                                                           [validation_set.__len__(), 300],
Anthony Larcher's avatar
Anthony Larcher committed
896
897
898
899
900
901
902
                                                           speaker_number,
                                                           model_archi)

    logging.critical(f"***{time.strftime('%H:%M:%S', time.localtime())} Initial metrics - Cross validation accuracy = {val_acc} %, EER = {val_eer * 100} %")
    logging.critical(f"***{time.strftime('%H:%M:%S', time.localtime())} Initial metrics - Test EER = {test_eer * 100} %")


Anthony Larcher's avatar
Anthony Larcher committed
903
    for epoch in range(1, epochs + 1):
904
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
905
906
907
        if curr_patience == 0:
            print(f"Stopping at epoch {epoch} for cause of patience")
            break
Anthony Larcher's avatar
Anthony Larcher committed
908
909
910
911
912
913
        model = train_epoch(model,
                            epoch,
                            training_loader,
                            optimizer,
                            dataset_params["log_interval"],
                            device=device,
Anthony Larcher's avatar
Anthony Larcher committed
914
915
                            clipping=clipping,
                            tb_writer=writer)
916
917

        # Add the cross validation here
Anthony Larcher's avatar
Anthony Larcher committed
918
919
920
        val_acc, val_loss, val_eer, test_eer = compute_metrics(model,
                                                               validation_loader,
                                                               device,
Anthony Larcher's avatar
Anthony Larcher committed
921
                                                               [validation_set.__len__(), 300],
Anthony Larcher's avatar
Anthony Larcher committed
922
923
924
925
926
                                                               speaker_number,
                                                               model_archi)

        logging.critical(f"***{time.strftime('%H:%M:%S', time.localtime())} Training metrics - Cross validation accuracy = {val_acc} %, EER = {val_eer * 100} %")
        logging.critical(f"***{time.strftime('%H:%M:%S', time.localtime())} Training metrics - Test EER = {test_eer * 100} %")
927
928
929
930

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

Anthony Larcher's avatar
Anthony Larcher committed
931
        # remember best accuracy and save checkpoint
Anthony Larcher's avatar
Anthony Larcher committed
932
933
        is_best = val_acc > best_accuracy
        best_accuracy = max(val_acc, best_accuracy)
Anthony Larcher's avatar
Anthony Larcher committed
934

Anthony Larcher's avatar
Anthony Larcher committed
935
936
937
938
939
940
        if type(model) is Xtractor:
            save_checkpoint({
                'epoch': epoch,
                'model_state_dict': model.state_dict(),
                'optimizer_state_dict': optimizer.state_dict(),
                'accuracy': best_accuracy,
Anthony Larcher's avatar
Anthony Larcher committed
941
942
                'scheduler': scheduler,
                'speaker_number' : speaker_number,
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
943
                'model_archi': model_archi
Anthony Larcher's avatar
Anthony Larcher committed
944
945
946
947
948
949
950
            }, 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,
Anthony Larcher's avatar
Anthony Larcher committed
951
952
                'scheduler': scheduler,
                'speaker_number': speaker_number,
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
953
                'model_archi': model_archi
Anthony Larcher's avatar
Anthony Larcher committed
954
            }, is_best, filename=tmp_model_name+".pt", best_filename=best_model_name+'.pt')
Anthony Larcher's avatar
Anthony Larcher committed
955
956
957

        if is_best:
            best_accuracy_epoch = epoch
Anthony Larcher's avatar
Anthony Larcher committed
958
959
960
            curr_patience = patience
        else:
            curr_patience -= 1
Anthony Larcher's avatar
Anthony Larcher committed
961
    #writer.close()
962

963
964
965
    for ii in range(torch.cuda.device_count()):
        print(torch.cuda.memory_summary(ii))

Anthony Larcher's avatar
Anthony Larcher committed
966
    logging.critical(f"Best accuracy {best_accuracy * 100.} obtained at epoch {best_accuracy_epoch}")
967

Anthony Larcher's avatar
Anthony Larcher committed
968
def train_epoch(model, epoch, training_loader, optimizer, log_interval, device, clipping=False, tb_writer=None):
969
970
971
972
    """

    :param model:
    :param epoch:
Anthony Larcher's avatar
Anthony Larcher committed
973
    :param training_loader:
974
    :param optimizer:
Anthony Larcher's avatar
Anthony Larcher committed
975
976
977
    :param log_interval:
    :param device:
    :param clipping:
978
979
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
980
    model.train()
Anthony Larcher's avatar
Anthony Larcher committed
981
    criterion = torch.nn.CrossEntropyLoss(reduction='mean')
982

Anthony Larcher's avatar
Anthony Larcher committed
983
984
985
986
987
    if isinstance(model, Xtractor):
        loss_criteria = model.loss
    else:
        loss_criteria = model.module.loss

988
    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
989
    running_loss = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
990
    for batch_idx, (data, target) in enumerate(training_loader):
Anthony Larcher's avatar
debug    
Anthony Larcher committed
991
        data = data.squeeze().to(device)
992
        target = target.squeeze()
Anthony Larcher's avatar
Anthony Larcher committed
993
        target = target.to(device)
994
        optimizer.zero_grad()
Anthony Larcher's avatar
Anthony Larcher committed
995
996

        if loss_criteria == 'aam':
Anthony Larcher's avatar
Anthony Larcher committed
997
            output, _ = model(data, target=target)
Anthony Larcher's avatar
Anthony Larcher committed
998
        else:
Anthony Larcher's avatar
debug    
Anthony Larcher committed
999
            output = model(data, target=None)
Anthony Larcher's avatar
Anthony Larcher committed
1000