xvector.py 53.9 KB
Newer Older
Anthony Larcher's avatar
Anthony Larcher committed
1
# coding: utf-8 -*-
Anthony Larcher's avatar
Anthony Larcher committed
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
#
# This file is part of SIDEKIT.
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is a python package for speaker verification.
# Home page: http://www-lium.univ-lemans.fr/sidekit/
#
# SIDEKIT is free software: you can redistribute it and/or modify
# it under the terms of the GNU LLesser General Public License as
# published by the Free Software Foundation, either version 3 of the License,
# or (at your option) any later version.
#
# SIDEKIT is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU Lesser General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with SIDEKIT.  If not, see <http://www.gnu.org/licenses/>.

"""
Anthony Larcher's avatar
Anthony Larcher committed
25
Copyright 2014-2020 Yevhenii Prokopalo, Anthony Larcher
Anthony Larcher's avatar
Anthony Larcher committed
26
"""
Anthony Larcher's avatar
Anthony Larcher committed
27

Anthony Larcher's avatar
Anthony Larcher committed
28
29
30

import pdb
import traceback
Anthony Larcher's avatar
Anthony Larcher committed
31
import logging
Anthony Larcher's avatar
Anthony Larcher committed
32
import matplotlib.pyplot as plt
Anthony Larcher's avatar
Anthony Larcher committed
33
import multiprocessing
Anthony Larcher's avatar
Anthony Larcher committed
34
import numpy
Anthony Larcher's avatar
Anthony Larcher committed
35
import pandas
Anthony Larcher's avatar
minor    
Anthony Larcher committed
36
import pickle
Anthony Larcher's avatar
Anthony Larcher committed
37
import shutil
38
import time
Anthony Larcher's avatar
Anthony Larcher committed
39
import torch
Anthony Larcher's avatar
Anthony Larcher committed
40
import torch.optim as optim
Anthony Larcher's avatar
Anthony Larcher committed
41
42
import yaml

Anthony Larcher's avatar
Anthony Larcher committed
43
from torchvision import transforms
Anthony Larcher's avatar
Anthony Larcher committed
44
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
45
from .xsets import SideSet
Anthony Larcher's avatar
Anthony Larcher committed
46
from .xsets import IdMapSet
Anthony Larcher's avatar
Anthony Larcher committed
47
from .res_net import RawPreprocessor, ResBlockWFMS
Anthony Larcher's avatar
Anthony Larcher committed
48
49
from ..bosaris import IdMap
from ..statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
50
from torch.utils.data import DataLoader
Anthony Larcher's avatar
Anthony Larcher committed
51
from sklearn.model_selection import train_test_split
Anthony Larcher's avatar
Anthony Larcher committed
52
from .sincnet import SincNet
Anthony Larcher's avatar
Anthony Larcher committed
53
#from torch.utils.tensorboard import SummaryWriter
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
54
55
56
from .loss import ArcLinear
from .loss import ArcFace
from .loss import l2_norm
Anthony Larcher's avatar
Anthony Larcher committed
57

Anthony Larcher's avatar
debug    
Anthony Larcher committed
58
59
from torch.nn.parallel import DistributedDataParallel as DDP

Anthony Larcher's avatar
Anthony Larcher committed
60
import tqdm
Anthony Larcher's avatar
Anthony Larcher committed
61

Anthony Larcher's avatar
Anthony Larcher committed
62
63
__license__ = "LGPL"
__author__ = "Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
64
__copyright__ = "Copyright 2015-2020 Anthony Larcher"
Anthony Larcher's avatar
Anthony Larcher committed
65
66
67
68
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
69
70


Anthony Larcher's avatar
Anthony Larcher committed
71
72
logging.basicConfig(format='%(asctime)s %(message)s')

Anthony Larcher's avatar
Anthony Larcher committed
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141

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


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

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

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

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









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

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


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

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


def plot_classes_preds(model, speech, labels):
    '''
Anthony Larcher's avatar
Anthony Larcher committed
142

Anthony Larcher's avatar
Anthony Larcher committed
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
    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



164
def get_lr(optimizer):
Anthony Larcher's avatar
Anthony Larcher committed
165
166
167
168
169
    """

    :param optimizer:
    :return:
    """
170
171
172
173
    for param_group in optimizer.param_groups:
        return param_group['lr']


Anthony Larcher's avatar
Anthony Larcher committed
174
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar'):
Anthony Larcher's avatar
Anthony Larcher committed
175
176
177
178
179
180
181
182
    """

    :param state:
    :param is_best:
    :param filename:
    :param best_filename:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
183
184
185
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, best_filename)
Anthony Larcher's avatar
Anthony Larcher committed
186

Anthony Larcher's avatar
Anthony Larcher committed
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
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
207

Anthony Larcher's avatar
Anthony Larcher committed
208

Anthony Larcher's avatar
Anthony Larcher committed
209
210
211
212
213
214
215
216
217
218
219
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
220
        super(GruPooling, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
        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
243

Anthony Larcher's avatar
Anthony Larcher committed
244
class Xtractor(torch.nn.Module):
245
246
247
    """
    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
248

Anthony Larcher's avatar
Anthony Larcher committed
249
250
251
    def __init__(self,
                 speaker_number,
                 model_archi="xvector",
Anthony Larcher's avatar
Anthony Larcher committed
252
                 loss=None,
Anthony Larcher's avatar
Anthony Larcher committed
253
254
255
                 norm_embedding=False,
                 aam_margin=0.5,
                 aam_s=0.5):
Anthony Larcher's avatar
Anthony Larcher committed
256
257
        """
        If config is None, default architecture is created
Anthony Larcher's avatar
Anthony Larcher committed
258
        :param model_archi:
Anthony Larcher's avatar
Anthony Larcher committed
259
        """
Anthony Larcher's avatar
Anthony Larcher committed
260
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
261
        self.speaker_number = speaker_number
Anthony Larcher's avatar
Anthony Larcher committed
262
        self.feature_size = None
Anthony Larcher's avatar
Anthony Larcher committed
263
        self.norm_embedding = norm_embedding
Anthony Larcher's avatar
Anthony Larcher committed
264

Anthony Larcher's avatar
Anthony Larcher committed
265
        if model_archi == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
266
267
268
269
270
271

            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
272
            self.feature_size = 30
Anthony Larcher's avatar
Anthony Larcher committed
273
274
275
            self.activation = torch.nn.LeakyReLU(0.2)

            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
276

Anthony Larcher's avatar
xv    
Anthony Larcher committed
277
            self.sequence_network = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
278
                ("conv1", torch.nn.Conv1d(self.feature_size, 512, 5, dilation=1)),
Anthony Larcher's avatar
Anthony Larcher committed
279
                ("activation1", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
280
                ("batch_norm1", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
281
282
                ("conv2", torch.nn.Conv1d(512, 512, 3, dilation=2)),
                ("activation2", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
283
                ("batch_norm2", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
284
285
                ("conv3", torch.nn.Conv1d(512, 512, 3, dilation=3)),
                ("activation3", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
286
                ("batch_norm3", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
287
                ("conv4", torch.nn.Conv1d(512, 512, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
288
                ("activation4", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
289
                ("batch_norm4", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
290
                ("conv5", torch.nn.Conv1d(512, 1536, 1)),
Anthony Larcher's avatar
Anthony Larcher committed
291
                ("activation5", torch.nn.LeakyReLU(0.2)),
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
292
                ("batch_norm5", torch.nn.BatchNorm1d(1536))
Anthony Larcher's avatar
Anthony Larcher committed
293
294
            ]))

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
295
296
            self.stat_pooling = MeanStdPooling()

Anthony Larcher's avatar
xv    
Anthony Larcher committed
297
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
298
                ("linear6", torch.nn.Linear(3072, 512))
Anthony Larcher's avatar
Anthony Larcher committed
299
300
            ]))

Anthony Larcher's avatar
Anthony Larcher committed
301
302
            if self.loss == "aam":
                self.after_speaker_embedding = torch.nn.Sequential(OrderedDict([
Anthony Larcher's avatar
Anthony Larcher committed
303
                  ("arclinear", ArcLinear(512, int(self.speaker_number), margin=aam_margin, s=aam_s))
Anthony Larcher's avatar
Anthony Larcher committed
304
305
306
307
                ]))
            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
308
                    ("batch_norm6", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
309
310
311
                    ("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
312
                    ("batch_norm7", torch.nn.BatchNorm1d(512)),
Anthony Larcher's avatar
Anthony Larcher committed
313
314
                    ("linear8", torch.nn.Linear(512, int(self.speaker_number)))
                ]))
Anthony Larcher's avatar
Anthony Larcher committed
315

Anthony Larcher's avatar
Anthony Larcher committed
316
317
318
319
            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
320
        elif model_archi == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
321
322
323
324
325
326

            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
327
328
329
            filts = [128, [128, 128], [128, 256], [256, 256]]
            self.norm_embedding = True

Anthony Larcher's avatar
Anthony Larcher committed
330
            self.preprocessor = RawPreprocessor(nb_samp=48000,
Anthony Larcher's avatar
Anthony Larcher committed
331
                                                in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
332
333
                                                out_channels=filts[0],
                                                kernel_size=3)
Anthony Larcher's avatar
Anthony Larcher committed
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350

            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
351
352
353
354
355
            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
356
            elif self.loss == "cce":
Anthony Larcher's avatar
Anthony Larcher committed
357
358
359
                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
360

Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
361
362
363
364
365
366
            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
367
        else:
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
368
369
370
371
372
373
            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
374

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
375
            self.loss = cfg["training"]["loss"]
Anthony Larcher's avatar
Anthony Larcher committed
376
            if self.loss == "aam":
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
377
378
                self.aam_margin = cfg["training"]["aam_margin"]
                self.aam_s = cfg["training"]["aam_s"]
Anthony Larcher's avatar
Anthony Larcher committed
379

Anthony Larcher's avatar
Anthony Larcher committed
380
381
382
            """
            Prepare Preprocessor
            """
Anthony Larcher's avatar
Anthony Larcher committed
383
            self.preprocessor = None
Anthony Larcher's avatar
Anthony Larcher committed
384
385
            if "preprocessor" in cfg:
                if cfg['preprocessor']["type"] == "sincnet":
Anthony Larcher's avatar
Anthony Larcher committed
386
                    self.preprocessor = SincNet(
Anthony Larcher's avatar
Anthony Larcher committed
387
388
389
390
391
392
393
394
395
396
397
398
                        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
399
                    self.feature_size = self.preprocessor.dimension
Anthony Larcher's avatar
Anthony Larcher committed
400
                elif cfg['preprocessor']["type"] == "rawnet2":
Anthony Larcher's avatar
Anthony Larcher committed
401
                    self.preprocessor = RawPreprocessor(nb_samp=int(cfg['preprocessor']["sampling_frequency"] * cfg['preprocessor']["duration"]),
Anthony Larcher's avatar
Anthony Larcher committed
402
                                                        in_channels=1,
Anthony Larcher's avatar
Anthony Larcher committed
403
404
405
406
407
408
                                                        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
409
                self.preprocessor_weight_decay = 0.000
Anthony Larcher's avatar
Anthony Larcher committed
410
411

            """
Anthony Larcher's avatar
minor    
Anthony Larcher committed
412
            Prepare sequence network
Anthony Larcher's avatar
Anthony Larcher committed
413
            """
Anthony Larcher's avatar
Anthony Larcher committed
414
            # Get Feature size
Anthony Larcher's avatar
Anthony Larcher committed
415
416
417
            if self.feature_size is None:
                self.feature_size = cfg["feature_size"]

Anthony Larcher's avatar
Anthony Larcher committed
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
            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
433
434
435
436
437
438
                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
439
                    segmental_layers.append((k, torch.nn.Conv1d(input_size,
Anthony Larcher's avatar
Anthony Larcher committed
440
                                                                cfg["segmental"][k]["output_channels"],
Anthony Larcher's avatar
Anthony Larcher committed
441
442
                                                                kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
443
444
                    input_size = cfg["segmental"][k]["output_channels"]

Anthony Larcher's avatar
Anthony Larcher committed
445
446
                elif k.startswith("ctrans"):
                    segmental_layers.append((k, torch.nn.ConvTranspose1d(input_size,
Anthony Larcher's avatar
debug    
Anthony Larcher committed
447
                                                                         cfg["segmental"][k][":"],
Anthony Larcher's avatar
Anthony Larcher committed
448
449
                                                                         kernel_size=cfg["segmental"][k]["kernel_size"],
                                                                         dilation=cfg["segmental"][k]["dilation"])))
Anthony Larcher's avatar
Anthony Larcher committed
450
451
452
                elif k.startswith("activation"):
                    segmental_layers.append((k, self.activation))

Anthony Larcher's avatar
Anthony Larcher committed
453
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
Anthony Larcher committed
454
455
                    segmental_layers.append((k, torch.nn.BatchNorm1d(input_size)))

Anthony Larcher's avatar
Anthony Larcher committed
456
            self.sequence_network = torch.nn.Sequential(OrderedDict(segmental_layers))
Anthony Larcher's avatar
Anthony Larcher committed
457
            self.sequence_network_weight_decay = cfg["segmental"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
458

Anthony Larcher's avatar
pooling    
Anthony Larcher committed
459
460
461
462
            """
            Pooling
            """
            self.stat_pooling = MeanStdPooling()
Anthony Larcher's avatar
Anthony Larcher committed
463
            tmp_input_size = input_size * 2
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
464
465
466
467
            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
468
                tmp_input_size = cfg["stat_pooling"]["gru_node"]
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
469

Anthony Larcher's avatar
Anthony Larcher committed
470
471
            self.stat_pooling_weight_decay = cfg["stat_pooling"]["weight_decay"]

Anthony Larcher's avatar
Anthony Larcher committed
472
            """
Anthony Larcher's avatar
Anthony Larcher committed
473
            Prepare last part of the network (after pooling)
Anthony Larcher's avatar
Anthony Larcher committed
474
            """
Anthony Larcher's avatar
Anthony Larcher committed
475
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
476
            input_size = tmp_input_size
Anthony Larcher's avatar
xv    
Anthony Larcher committed
477
478
            before_embedding_layers = []
            for k in cfg["before_embedding"].keys():
Anthony Larcher's avatar
Anthony Larcher committed
479
                if k.startswith("lin"):
Anthony Larcher's avatar
Anthony Larcher committed
480
481
                    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
482
                    else:
Anthony Larcher's avatar
Anthony Larcher committed
483
484
                        before_embedding_layers.append((k, torch.nn.Linear(input_size,
                                                                           cfg["before_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
485
                        input_size = cfg["before_embedding"][k]["output"]
Anthony Larcher's avatar
Anthony Larcher committed
486
487

                elif k.startswith("activation"):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
488
                    before_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
Anthony Larcher committed
489

Anthony Larcher's avatar
Anthony Larcher committed
490
                elif k.startswith('batch_norm'):
Anthony Larcher's avatar
xv    
Anthony Larcher committed
491
                    before_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
Anthony Larcher committed
492
493

                elif k.startswith('dropout'):
Anthony Larcher's avatar
Anthony Larcher committed
494
                    before_embedding_layers.append((k, torch.nn.Dropout(p=cfg["before_embedding"][k])))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
495

Anthony Larcher's avatar
Anthony Larcher committed
496
            self.before_speaker_embedding = torch.nn.Sequential(OrderedDict(before_embedding_layers))
Anthony Larcher's avatar
Anthony Larcher committed
497
            self.before_speaker_embedding_weight_decay = cfg["before_embedding"]["weight_decay"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
498

Anthony Larcher's avatar
Anthony Larcher committed
499
            # if loss_criteria is "cce"
Anthony Larcher's avatar
xv    
Anthony Larcher committed
500
            # Create sequential object for the second part of the network
Anthony Larcher's avatar
Anthony Larcher committed
501
502
503
504
505
506
507
508
            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
509
                                                                          cfg["after_embedding"][k]["output"])))
Anthony Larcher's avatar
Anthony Larcher committed
510
                            input_size = cfg["after_embedding"][k]["output"]
Anthony Larcher's avatar
xv    
Anthony Larcher committed
511

Anthony Larcher's avatar
Anthony Larcher committed
512
513
514
515
516
                    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
517

Anthony Larcher's avatar
Anthony Larcher committed
518
519
                    elif k.startswith("activation"):
                        after_embedding_layers.append((k, self.activation))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
520

Anthony Larcher's avatar
Anthony Larcher committed
521
522
                    elif k.startswith('batch_norm'):
                        after_embedding_layers.append((k, torch.nn.BatchNorm1d(input_size)))
Anthony Larcher's avatar
xv    
Anthony Larcher committed
523

Anthony Larcher's avatar
Anthony Larcher committed
524
525
526
527
528
529
                    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
530
                self.norm_embedding = True
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
531
532
533
534
535
536
537
                #self.after_speaker_embedding = ArcLinear(input_size,
                #                                         self.speaker_number,
                #                                         margin=self.aam_margin,
                #                                         s=self.aam_s)
                self.after_speaker_embedding = ArcFace(embedding_size=input_size,
                                                       classnum=self.speaker_number,
                                                       s=64.,
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
538
                                                       m=0.5)
Anthony Larcher's avatar
Anthony Larcher committed
539

Anthony Larcher's avatar
Anthony Larcher committed
540
            self.after_speaker_embedding_weight_decay = cfg["after_embedding"]["weight_decay"]
Anthony Larcher's avatar
Anthony Larcher committed
541

Anthony Larcher's avatar
Anthony Larcher committed
542

Anthony Larcher's avatar
Anthony Larcher committed
543
    def forward(self, x, is_eval=False, target=None):
544
545
546
        """

        :param x:
Anthony Larcher's avatar
Anthony Larcher committed
547
        :param is_eval:
548
549
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
550
551
552
        if self.preprocessor is not None:
            x = self.preprocessor(x)

Anthony Larcher's avatar
Anthony Larcher committed
553
        x = self.sequence_network(x)
554

Anthony Larcher's avatar
Anthony Larcher committed
555
        # Mean and Standard deviation pooling
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
556
        x = self.stat_pooling(x)
Anthony Larcher's avatar
Anthony Larcher committed
557
558

        x = self.before_speaker_embedding(x)
559

Anthony Larcher's avatar
Anthony Larcher committed
560
        if self.norm_embedding:
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
561
            #x_norm = x.norm(p=2,dim=1, keepdim=True) / 10. # Why  10. ?
Anthony Larcher's avatar
arcface    
Anthony Larcher committed
562
563
564
            #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
565

Anthony Larcher's avatar
Anthony Larcher committed
566
567
568
        if is_eval:
            return x

Anthony Larcher's avatar
Anthony Larcher committed
569
570
        if self.loss == "cce":
            x = self.after_speaker_embedding(x)
Anthony Larcher's avatar
Anthony Larcher committed
571

Anthony Larcher's avatar
Anthony Larcher committed
572
573
574
575
576
577
        elif self.loss == "aam":
            if not is_eval:
                x = self.after_speaker_embedding(x,target=target)
            else:
                x = self.after_speaker_embedding(x, target=None)

Anthony Larcher's avatar
Anthony Larcher committed
578
        return x
Anthony Larcher's avatar
Anthony Larcher committed
579

580
581
    def context_size(self):
        context = 1
Anthony Larcher's avatar
Anthony Larcher committed
582
583
584
585
586
587
588
589
590
        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
591

Anthony Larcher's avatar
Anthony Larcher committed
592
def xtrain(speaker_number,
Anthony Larcher's avatar
Anthony Larcher committed
593
           dataset_yaml,
Anthony Larcher's avatar
Anthony Larcher committed
594
595
           epochs=None,
           lr=None,
Anthony Larcher's avatar
Anthony Larcher committed
596
           model_yaml=None,
Anthony Larcher's avatar
Anthony Larcher committed
597
           model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
598
599
600
601
           loss=None,
           aam_margin=None,
           aam_s=None,
           patience=None,
Anthony Larcher's avatar
Anthony Larcher committed
602
           tmp_model_name=None,
Anthony Larcher's avatar
minor    
Anthony Larcher committed
603
           best_model_name=None,
Anthony Larcher's avatar
Anthony Larcher committed
604
           multi_gpu=True,
605
           clipping=False,
Anthony Larcher's avatar
Anthony Larcher committed
606
           opt=None,
Anthony Larcher's avatar
Anthony Larcher committed
607
608
           reset_parts=[],
           freeze_parts=[],
Anthony Larcher's avatar
Anthony Larcher committed
609
           num_thread=None):
610
611
    """

Anthony Larcher's avatar
Anthony Larcher committed
612
613
614
615
616
617
    :param speaker_number:
    :param dataset_yaml:
    :param epochs:
    :param lr:
    :param model_yaml:
    :param model_name:
Anthony Larcher's avatar
Anthony Larcher committed
618
619
620
621
    :param loss:
    :param aam_margin:
    :param aam_s:
    :param patience:
Anthony Larcher's avatar
Anthony Larcher committed
622
623
624
625
    :param tmp_model_name:
    :param best_model_name:
    :param multi_gpu:
    :param clipping:
Anthony Larcher's avatar
Anthony Larcher committed
626
627
628
    :param opt:
    :param reset_parts:
    :param freeze_parts:
Anthony Larcher's avatar
Anthony Larcher committed
629
    :param num_thread:
630
631
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
632
633
634
635
    # 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
636

Anthony Larcher's avatar
Anthony Larcher committed
637
638
639
    if num_thread is None:
        num_thread = multiprocessing.cpu_count()

Anthony Larcher's avatar
Anthony Larcher committed
640
    logging.critical(f"Start process at {time.strftime('%H:%M:%S', time.localtime())}")
641

Anthony Larcher's avatar
Anthony Larcher committed
642
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
Anthony Larcher's avatar
Anthony Larcher committed
643

Anthony Larcher's avatar
Anthony Larcher committed
644
    # Start from scratch
Anthony Larcher's avatar
Anthony Larcher committed
645
    if model_name is None and model_yaml in ["xvector", "rawnet2"]:
Anthony Larcher's avatar
Anthony Larcher committed
646
        # Initialize a first model
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
647
        if model_yaml == "xvector":
Anthony Larcher's avatar
Anthony Larcher committed
648
            model = Xtractor(speaker_number, "xvector")
Anthony Larcher's avatar
rawnet2    
Anthony Larcher committed
649
        elif model_yaml == "rawnet2":
Anthony Larcher's avatar
pooling    
Anthony Larcher committed
650
            model = Xtractor(speaker_number, "rawnet2")
Anthony Larcher's avatar
Anthony Larcher committed
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
    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
677
        if model_name is None:
Anthony Larcher's avatar
Anthony Larcher committed
678
679
680
            model = Xtractor(speaker_number, model_yaml)

         # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
681
        else:
Anthony Larcher's avatar
Anthony Larcher committed
682
683
684
            # Load the model
            logging.critical(f"*** Load model from = {model_name}")
            checkpoint = torch.load(model_name)
Anthony Larcher's avatar
Anthony Larcher committed
685
            model = Xtractor(speaker_number, model_yaml)
Anthony Larcher's avatar
Anthony Larcher committed
686

Anthony Larcher's avatar
fix API    
Anthony Larcher committed
687
688
            """
            Here we remove all layers that we don't want to reload
Anthony Larcher's avatar
Anthony Larcher committed
689
        
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
690
691
692
693
694
695
696
697
698
699
700
701
702
            """
            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
703

Anthony Larcher's avatar
Anthony Larcher committed
704
    print(model)
Anthony Larcher's avatar
Anthony Larcher committed
705

Anthony Larcher's avatar
Anthony Larcher committed
706
    if torch.cuda.device_count() > 1 and multi_gpu:
Anthony Larcher's avatar
Anthony Larcher committed
707
708
        print("Let's use", torch.cuda.device_count(), "GPUs!")
        model = torch.nn.DataParallel(model)
Anthony Larcher's avatar
Anthony Larcher committed
709
        #model = DDP(model)
Anthony Larcher's avatar
Anthony Larcher committed
710
711
    else:
        print("Train on a single GPU")
Anthony Larcher's avatar
Anthony Larcher committed
712
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
713
714

    """
Anthony Larcher's avatar
Anthony Larcher committed
715
716
717
718
    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
719
    """
Anthony Larcher's avatar
Anthony Larcher committed
720
721
722
723
    with open(dataset_yaml, "r") as fh:
        dataset_params = yaml.load(fh, Loader=yaml.FullLoader)
        df = pandas.read_csv(dataset_params["dataset_description"])
    training_df, validation_df = train_test_split(df, test_size=dataset_params["validation_ratio"])
Anthony Larcher's avatar
Anthony Larcher committed
724
    torch.manual_seed(dataset_params['seed'])
725
726
727
    training_set = SideSet(dataset_yaml, 
                           set_type="train", 
                           dataset_df=training_df, 
Anthony Larcher's avatar
Anthony Larcher committed
728
729
                           chunk_per_segment=dataset_params['train']['chunk_per_segment'], 
                           overlap=dataset_params['train']['overlap'])
Anthony Larcher's avatar
Anthony Larcher committed
730
731
732
    training_loader = DataLoader(training_set,
                                 batch_size=dataset_params["batch_size"],
                                 shuffle=True,
Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
733
                                 drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
734
                                 pin_memory=True,
Anthony Larcher's avatar
Anthony Larcher committed
735
                                 num_workers=num_thread)
736

Anthony Larcher's avatar
Anthony Larcher committed
737
738
739
    validation_set = SideSet(dataset_yaml, set_type="validation", dataset_df=validation_df)
    validation_loader = DataLoader(validation_set,
                                   batch_size=dataset_params["batch_size"],
Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
740
                                   drop_last=True,
Anthony Larcher's avatar
Anthony Larcher committed
741
                                   pin_memory=True,
Anthony Larcher's avatar
Anthony Larcher committed
742
                                   num_workers=num_thread)
Anthony Larcher's avatar
Anthony Larcher committed
743

Anthony Larcher's avatar
Anthony Larcher committed
744
745
746
747
748
749
    # Add for TensorBoard
    #dataiter = iter(training_loader)
    #data, labels = dataiter.next()
    #writer.add_graph(model, data)


Anthony Larcher's avatar
Anthony Larcher committed
750
751
752
    """
    Set the training options
    """
Anthony Larcher's avatar
Anthony Larcher committed
753
    if opt == 'adam':
Anthony Larcher's avatar
Anthony Larcher committed
754
        _optimizer = torch.optim.Adam
Anthony Larcher's avatar
Anthony Larcher committed
755
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
756
757
    elif opt == 'rmsprop':
        _optimizer = torch.optim.RMSprop
Anthony Larcher's avatar
Anthony Larcher committed
758
        _options = {'lr': lr}
Anthony Larcher's avatar
Anthony Larcher committed
759
760
761
    else: # opt == 'sgd'
        _optimizer = torch.optim.SGD
        _options = {'lr': lr, 'momentum': 0.9}
Anthony Larcher's avatar
Anthony Larcher committed
762

Anthony Larcher's avatar
Anthony Larcher committed
763
764
765
766
767
768
769
770
771
772
773
774
775
776
    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
777
    param_list = []
Anthony Larcher's avatar
Anthony Larcher committed
778
    if type(model) is Xtractor:
Anthony Larcher's avatar
Anthony Larcher committed
779
780
781
782
783
784
785
        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
786
    else:
Anthony Larcher's avatar
Anthony Larcher committed
787
788
789
790
791
792
793
794
795
        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
796
797
798
799
800
801

    #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
802

Anthony Larcher's avatar
Anthony Larcher committed
803
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', verbose=True)
804

Anthony Larcher's avatar
sincxv    
Anthony Larcher committed
805
    best_accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
806
    best_accuracy_epoch = 1
Anthony Larcher's avatar
Anthony Larcher committed
807
    curr_patience = patience
Anthony Larcher's avatar
Anthony Larcher committed
808
    for epoch in range(1, epochs + 1):
809
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
810
811
812
        if curr_patience == 0:
            print(f"Stopping at epoch {epoch} for cause of patience")
            break
Anthony Larcher's avatar
Anthony Larcher committed
813
814
815
816
817
818
        model = train_epoch(model,
                            epoch,
                            training_loader,
                            optimizer,
                            dataset_params["log_interval"],
                            device=device,
Anthony Larcher's avatar
Anthony Larcher committed
819
820
                            clipping=clipping,
                            tb_writer=writer)
821
822

        # Add the cross validation here
Anthony Larcher's avatar
Anthony Larcher committed
823
        accuracy, val_loss = cross_validation(model, validation_loader, device=device)
Anthony Larcher's avatar
Anthony Larcher committed
824
        logging.critical(f"***{time.strftime('%H:%M:%S', time.localtime())} Cross validation accuracy = {accuracy} %")
825
826
827
828

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

Anthony Larcher's avatar
Anthony Larcher committed
829
830
831
832
        # remember best accuracy and save checkpoint
        is_best = accuracy > best_accuracy
        best_accuracy = max(accuracy, best_accuracy)

Anthony Larcher's avatar
Anthony Larcher committed
833
834
835
836
837
838
        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
839
840
                'scheduler': scheduler,
                'speaker_number' : speaker_number,
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
841
                'model_archi': model_archi
Anthony Larcher's avatar
Anthony Larcher committed
842
843
844
845
846
847
848
            }, 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
849
850
                'scheduler': scheduler,
                'speaker_number': speaker_number,
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
851
                'model_archi': model_archi
Anthony Larcher's avatar
Anthony Larcher committed
852
            }, is_best, filename=tmp_model_name+".pt", best_filename=best_model_name+'.pt')
Anthony Larcher's avatar
Anthony Larcher committed
853
854
855

        if is_best:
            best_accuracy_epoch = epoch
Anthony Larcher's avatar
Anthony Larcher committed
856
857
858
            curr_patience = patience
        else:
            curr_patience -= 1
Anthony Larcher's avatar
Anthony Larcher committed
859
    #writer.close()
860

861
862
863
    for ii in range(torch.cuda.device_count()):
        print(torch.cuda.memory_summary(ii))

Anthony Larcher's avatar
Anthony Larcher committed
864
    logging.critical(f"Best accuracy {best_accuracy * 100.} obtained at epoch {best_accuracy_epoch}")
865

Anthony Larcher's avatar
Anthony Larcher committed
866
def train_epoch(model, epoch, training_loader, optimizer, log_interval, device, clipping=False, tb_writer=None):
867
868
869
870
    """

    :param model:
    :param epoch:
Anthony Larcher's avatar
Anthony Larcher committed
871
    :param training_loader:
872
    :param optimizer:
Anthony Larcher's avatar
Anthony Larcher committed
873
874
875
    :param log_interval:
    :param device:
    :param clipping:
876
877
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
878
    model.train()
Anthony Larcher's avatar
Anthony Larcher committed
879
    criterion = torch.nn.CrossEntropyLoss(reduction='mean')
880

Anthony Larcher's avatar
Anthony Larcher committed
881
882
883
884
885
    if isinstance(model, Xtractor):
        loss_criteria = model.loss
    else:
        loss_criteria = model.module.loss

886
    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
887
    running_loss = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
888
    for batch_idx, (data, target) in enumerate(training_loader):
889
        target = target.squeeze()
Anthony Larcher's avatar
Anthony Larcher committed
890
        target = target.to(device)
891
        optimizer.zero_grad()
Anthony Larcher's avatar
Anthony Larcher committed
892
893
894
895
896
897

        if loss_criteria == 'aam':
            output = model(data.to(device), target=target)
        else:
            output = model(data.to(device), target=None)

Anthony Larcher's avatar
Anthony Larcher committed
898
        #with GuruMeditation():
Anthony Larcher's avatar
Anthony Larcher committed
899
        loss = criterion(output, target)
Anthony Larcher's avatar
Anthony Larcher committed
900
901
902
903
904
905
906
907
        if not torch.isnan(loss):
            loss.backward()
            if clipping:
                torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
            running_loss += loss.item()
            optimizer.step()

            running_loss += loss.item()
Anthony Larcher's avatar
Anthony Larcher committed
908
            accuracy += (torch.argmax(output.data, 1) == target).sum()
Anthony Larcher's avatar
Anthony Larcher committed
909
910
            if batch_idx % log_interval == 0:
                batch_size = target.shape[0]
Anthony Larcher's avatar
debug    
Anthony Larcher committed
911
912
                logging.critical('{}, Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
                    time.strftime('%H:%M:%S', time.localtime()),
Anthony Larcher's avatar
Anthony Larcher committed
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
                    epoch, batch_idx + 1, training_loader.__len__(),
                    100. * batch_idx / training_loader.__len__(), loss.item(),
                    100.0 * accuracy.item() / ((batch_idx + 1) * batch_size)))
                #tb_writer.add_scalar('training loss',
                #                     running_loss / log_interval,
                #                     epoch * len(training_loader) + batch_idx)
                #tb_writer.add_scalar('training_accuracy',
                #                      100.0 * accuracy.item() / ((batch_idx + 1) * batch_size),
                #                      epoch * len(training_loader) + batch_idx)

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

929
930
931
932
933
934
935
936
937
938
939
940
941
        else:
            save_checkpoint({
                             'epoch': epoch,
                             'model_state_dict': model.state_dict(),
                             'optimizer_state_dict': optimizer.state_dict(),
                             'accuracy': 0.0,
                             'scheduler': 0.0
                             }, False, filename="model_loss_NAN.pt", best_filename='toto.pt')
            with open("batch_loss_NAN.pkl", "wb") as fh:
                pickle.dump(data.cpu(), fh)
            import sys
            sys.exit()
        running_loss = 0.0
942
943
944
    return model


Anthony Larcher's avatar
Anthony Larcher committed
945
def cross_validation(model, validation_loader, device):
946
947
948
    """

    :param model:
Anthony Larcher's avatar
Anthony Larcher committed
949
950
    :param validation_loader:
    :param device:
951
952
953
954
    :return:
    """
    model.eval()

Anthony Larcher's avatar
Anthony Larcher committed
955
956
957
958
959
    if isinstance(model, Xtractor):
        loss_criteria = model.loss
    else:
        loss_criteria = model.module.loss

960
    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
961
    loss = 0.0
962
    criterion = torch.nn.CrossEntropyLoss()
Anthony Larcher's avatar
Anthony Larcher committed
963
964
965
966
    with torch.no_grad():
        for batch_idx, (data, target) in enumerate(validation_loader):
            batch_size = target.shape[0]
            target = target.squeeze()
Anthony Larcher's avatar
Anthony Larcher committed
967
968
969
970
971
972

            if loss_criteria == "aam":
                output = model(data.to(device), target=target)
            else:
                output = model(data.to(device), target=None)

Anthony Larcher's avatar
Anthony Larcher committed
973
            accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()
974

Anthony Larcher's avatar
Anthony Larcher committed
975
            loss += criterion(output, target.to(device))
Anthony Larcher's avatar
Anthony Larcher committed
976

Anthony Larcher's avatar
Anthony Larcher committed
977
978
979
980
    return 100. * accuracy.cpu().numpy() / ((batch_idx + 1) * batch_size), \
           loss.cpu().numpy() / ((batch_idx + 1) * batch_size)


Anthony Larcher's avatar
Anthony Larcher committed
981
982
def extract_embeddings(idmap_name,
                       model_filename,
Anthony Larcher's avatar
Anthony Larcher committed
983
                       data_root_name,
Anthony Larcher's avatar
Anthony Larcher committed
984
                       device,
Anthony Larcher's avatar
Anthony Larcher committed
985
986
                       model_yaml=None,
                       speaker_number=None,
Anthony Larcher's avatar
Anthony Larcher committed
987
                       file_extension="wav",
988
                       transform_pipeline=None,
989
990
                       frame_shift=0.01,
                       frame_duration=0.025,
991
                       num_thread=1):
992
993
994
    # Load the model
    if isinstance(model_filename, str):
        checkpoint = torch.load(model_filename)
Anthony Larcher's avatar
Anthony Larcher committed
995
996
997
        if speaker_number is None:
            speaker_number = checkpoint["speaker_number"]
        if model_yaml is None:
Anthony Larcher's avatar
fix API    
Anthony Larcher committed
998
999
            model_archi = checkpoint["model_archi"]
        model = Xtractor(speaker_number, model_archi=model_archi)
1000
        model.load_state_dict(checkpoint["model_state_dict"])