xvector.py 25.7 KB
Newer Older
Anthony Larcher's avatar
Anthony Larcher committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
# -*- coding: utf-8 -*-
#
# 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 2014-2019 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
import h5py
import logging
Anthony Larcher's avatar
Anthony Larcher committed
30
import sys
Anthony Larcher's avatar
Anthony Larcher committed
31
import numpy
Anthony Larcher's avatar
minor    
Anthony Larcher committed
32
import pickle
Anthony Larcher's avatar
Anthony Larcher committed
33
import torch
Anthony Larcher's avatar
Anthony Larcher committed
34
35
import torch.optim as optim
import torch.multiprocessing as mp
Anthony Larcher's avatar
Anthony Larcher committed
36
from torchvision import transforms
Anthony Larcher's avatar
Anthony Larcher committed
37
from collections import OrderedDict
Anthony Larcher's avatar
Anthony Larcher committed
38
from sidekit.nnet.xsets import XvectorMultiDataset, XvectorDataset, StatDataset, VoxDataset
Anthony Larcher's avatar
Anthony Larcher committed
39
from sidekit.nnet.xsets import FrequencyMask, CMVN, TemporalMask
Anthony Larcher's avatar
Anthony Larcher committed
40
41
from sidekit.bosaris import IdMap
from sidekit.statserver import StatServer
Anthony Larcher's avatar
Anthony Larcher committed
42
from torch.utils.data import DataLoader
Anthony Larcher's avatar
Anthony Larcher committed
43

Anthony Larcher's avatar
Anthony Larcher committed
44
45
46
47
48
49
50
__license__ = "LGPL"
__author__ = "Anthony Larcher"
__copyright__ = "Copyright 2015-2019 Anthony Larcher"
__maintainer__ = "Anthony Larcher"
__email__ = "anthony.larcher@univ-lemans.fr"
__status__ = "Production"
__docformat__ = 'reS'
Anthony Larcher's avatar
Anthony Larcher committed
51
52


Anthony Larcher's avatar
Anthony Larcher committed
53
#logging.basicConfig(stream=sys.stdout, level=logging.INFO)
Anthony Larcher's avatar
Anthony Larcher committed
54
55


Anthony Larcher's avatar
Anthony Larcher committed
56
57
58
59
60
61
62
def split_file_list(batch_files, num_processes):
    # Cut the list of files into args.num_processes lists of files
    batch_sub_lists = [[]] * num_processes
    x = [ii for ii in range(len(batch_files))]
    for ii in range(num_processes):
        batch_sub_lists[ii - 1] = [batch_files[z + ii] for z in x[::num_processes] if (z + ii) < len(batch_files)]
    return batch_sub_lists
Anthony Larcher's avatar
Anthony Larcher committed
63
64
65


class Xtractor(torch.nn.Module):
66
67
68
    """
    Class that defines an x-vector extractor based on 5 convolutional layers and a mean standard deviation pooling
    """
Anthony Larcher's avatar
Anthony Larcher committed
69
    def __init__(self, spk_number, dropout):
Anthony Larcher's avatar
Anthony Larcher committed
70
        super(Xtractor, self).__init__()
Anthony Larcher's avatar
Anthony Larcher committed
71
        self.frame_conv0 = torch.nn.Conv1d(30, 512, 5, dilation=1)
Anthony Larcher's avatar
Anthony Larcher committed
72
73
74
        self.frame_conv1 = torch.nn.Conv1d(512, 512, 3, dilation=2)
        self.frame_conv2 = torch.nn.Conv1d(512, 512, 3, dilation=3)
        self.frame_conv3 = torch.nn.Conv1d(512, 512, 1)
Anthony Larcher's avatar
test    
Anthony Larcher committed
75
76
        self.frame_conv4 = torch.nn.Conv1d(512, 3 * 512, 1)
        self.seg_lin0 = torch.nn.Linear(3 * 512 * 2, 512)
Anthony Larcher's avatar
Anthony Larcher committed
77
        self.dropout_lin0 = torch.nn.Dropout(p=dropout)
Anthony Larcher's avatar
Anthony Larcher committed
78
        self.seg_lin1 = torch.nn.Linear(512, 512)
Anthony Larcher's avatar
Anthony Larcher committed
79
        self.dropout_lin1 = torch.nn.Dropout(p=dropout)
Anthony Larcher's avatar
Anthony Larcher committed
80
81
82
        self.seg_lin2 = torch.nn.Linear(512, spk_number)
        #
        self.norm0 = torch.nn.BatchNorm1d(512)
Anthony Larcher's avatar
Anthony Larcher committed
83
84
85
        self.norm1 = torch.nn.BatchNorm1d(512)
        self.norm2 = torch.nn.BatchNorm1d(512)
        self.norm3 = torch.nn.BatchNorm1d(512)
Anthony Larcher's avatar
test    
Anthony Larcher committed
86
        self.norm4 = torch.nn.BatchNorm1d(3 * 512)
Anthony Larcher's avatar
Anthony Larcher committed
87
        self.norm6 = torch.nn.BatchNorm1d(512)
Anthony Larcher's avatar
Anthony Larcher committed
88
        self.norm7 = torch.nn.BatchNorm1d(512)
Anthony Larcher's avatar
Anthony Larcher committed
89
        #
Anthony Larcher's avatar
Anthony Larcher committed
90
        self.activation = torch.nn.LeakyReLU(0.2)
Anthony Larcher's avatar
Anthony Larcher committed
91

92
    def produce_embeddings(self, x):
Anthony Larcher's avatar
Anthony Larcher committed
93
        """
Anthony Larcher's avatar
Anthony Larcher committed
94

95
96
97
        :param x:
        :return:
        """
Anthony Larcher's avatar
Anthony Larcher committed
98
99
100
101
102
        frame_emb_0 = self.norm0(self.activation(self.frame_conv0(x)))
        frame_emb_1 = self.norm1(self.activation(self.frame_conv1(frame_emb_0)))
        frame_emb_2 = self.norm2(self.activation(self.frame_conv2(frame_emb_1)))
        frame_emb_3 = self.norm3(self.activation(self.frame_conv3(frame_emb_2)))
        frame_emb_4 = self.norm4(self.activation(self.frame_conv4(frame_emb_3)))
Anthony Larcher's avatar
Anthony Larcher committed
103
104
105

        mean = torch.mean(frame_emb_4, dim=2)
        std = torch.std(frame_emb_4, dim=2)
106
        seg_emb = torch.cat([mean, std], dim=1)
Anthony Larcher's avatar
Anthony Larcher committed
107

108
109
        embedding_a = self.seg_lin0(seg_emb)
        return embedding_a
Anthony Larcher's avatar
Anthony Larcher committed
110
111

    def forward(self, x):
112
113
114
115
116
117
        """

        :param x:
        :return:
        """
        seg_emb_0 = self.produce_embeddings(x)
Anthony Larcher's avatar
Anthony Larcher committed
118
        # batch-normalisation after this layer
119
        seg_emb_1 = self.norm6(self.activation(seg_emb_0))
Anthony Larcher's avatar
Anthony Larcher committed
120
        # new layer with batch Normalization
121
        seg_emb_2 = self.norm7(self.activation(self.seg_lin1(self.dropout_lin1(seg_emb_1))))
Anthony Larcher's avatar
Anthony Larcher committed
122
        # No batch-normalisation after this layer
123
        result = self.activation(self.seg_lin2(seg_emb_2))
Anthony Larcher's avatar
Anthony Larcher committed
124
125
        return result

126
127
128
129
130
131
132
133
134
135
136
    def extract(self, x):
        """
        Extract x-vector given an input sequence of features

        :param x:
        :return:
        """
        embedding_a = self.produce_embeddings(x)
        embedding_b = self.seg_lin1(self.norm6(self.activation(embedding_a)))

        return embedding_a, embedding_b
Anthony Larcher's avatar
Anthony Larcher committed
137
138
139

    def init_weights(self):
        """
140
        Initialize the x-vector extract weights and biaises
Anthony Larcher's avatar
Anthony Larcher committed
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
        """
        torch.nn.init.normal_(self.frame_conv0.weight, mean=-0.5, std=0.1)
        torch.nn.init.normal_(self.frame_conv1.weight, mean=-0.5, std=0.1)
        torch.nn.init.normal_(self.frame_conv2.weight, mean=-0.5, std=0.1)
        torch.nn.init.normal_(self.frame_conv3.weight, mean=-0.5, std=0.1)
        torch.nn.init.normal_(self.frame_conv4.weight, mean=-0.5, std=0.1)
        torch.nn.init.xavier_uniform(self.seg_lin0.weight)
        torch.nn.init.xavier_uniform(self.seg_lin1.weight)
        torch.nn.init.xavier_uniform(self.seg_lin2.weight)

        torch.nn.init.constant(self.frame_conv0.bias, 0.1)
        torch.nn.init.constant(self.frame_conv1.bias, 0.1)
        torch.nn.init.constant(self.frame_conv2.bias, 0.1)
        torch.nn.init.constant(self.frame_conv3.bias, 0.1)
        torch.nn.init.constant(self.frame_conv4.bias, 0.1)
        torch.nn.init.constant(self.seg_lin0.bias, 0.1)
        torch.nn.init.constant(self.seg_lin1.bias, 0.1)
        torch.nn.init.constant(self.seg_lin2.bias, 0.1)

Anthony Larcher's avatar
Anthony Larcher committed
160
161

def xtrain(args):
162
163
164
165
166
167
    """
    Initialize and train an x-vector in asynchronous manner

    :param args:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
168
    # Initialize a first model and save to disk
Anthony Larcher's avatar
Anthony Larcher committed
169
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
170
171
172
173
174
175
176
177
    current_model_file_name = "initial_model"
    torch.save(model.state_dict(), current_model_file_name)

    for epoch in range(1, args.epochs + 1):
        current_model_file_name = train_epoch(epoch, args, current_model_file_name)

        # Add the cross validation here
        accuracy = cross_validation(args, current_model_file_name)
Anthony Larcher's avatar
Anthony Larcher committed
178
        print("*** Cross validation accuracy = {} %".format(accuracy))
Anthony Larcher's avatar
Anthony Larcher committed
179

Anthony Larcher's avatar
Anthony Larcher committed
180
        # Decrease learning rate after every epoch
Anthony Larcher's avatar
Anthony Larcher committed
181
        #args.lr = args.lr * 0.9
Anthony Larcher's avatar
sad    
Anthony Larcher committed
182
183
        args.lr = args.lr * 0.9
        print("        Decrease learning rate: {}".format(args.lr))
Anthony Larcher's avatar
Anthony Larcher committed
184

Anthony Larcher's avatar
Anthony Larcher committed
185
186

def train_epoch(epoch, args, initial_model_file_name):
187
188
189
190
191
192
193
194
    """
    Process one training epoch using an asynchronous implementation of the training

    :param epoch:
    :param args:
    :param initial_model_file_name:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
    # Compute the megabatch number
    with open(args.batch_training_list, 'r') as fh:
        batch_file_list = [l.rstrip() for l in fh]

    # Shorten the batch_file_list to be a multiple of

    megabatch_number = len(batch_file_list) // (args.averaging_step * args.num_processes)
    megabatch_size = args.averaging_step * args.num_processes
    print("Epoch {}, number of megabatches = {}".format(epoch, megabatch_number))

    current_model = initial_model_file_name

    # For each sublist: run an asynchronous training and averaging of the model
    for ii in range(megabatch_number):
        print('Process megabatch [{} / {}]'.format(ii + 1, megabatch_number))
        current_model = train_asynchronous(epoch,
                                           args,
                                           current_model,
                                           batch_file_list[megabatch_size * ii: megabatch_size * (ii + 1)],
                                           ii,
                                           megabatch_number)  # function that split train, fuse and write the new model to disk
    return current_model


219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
def train_asynchronous(epoch, args, initial_model_file_name, batch_file_list, megabatch_idx, megabatch_number):
    """
    Process one mega-batch of data asynchronously, average the model parameters across
    subrocesses and return the updated version of the model

    :param epoch:
    :param args:
    :param initial_model_file_name:
    :param batch_file_list:
    :param megabatch_idx:
    :param megabatch_number:
    :return:
    """
    # Split the list of files for each process
    sub_lists = split_file_list(batch_file_list, args.num_processes)

    #
    output_queue = mp.Queue()
    # output_queue = multiprocessing.Queue()

    processes = []
    for rank in range(args.num_processes):
        p = mp.Process(target=train_worker,
                       args=(rank, epoch, args, initial_model_file_name, sub_lists[rank], output_queue)
                       )
        # We first train the model across `num_processes` processes
        p.start()
        processes.append(p)

    # Average the models and write the new one to disk
    asynchronous_model = []
    for ii in range(args.num_processes):
        asynchronous_model.append(dict(output_queue.get()))

    for p in processes:
        p.join()

    av_model = Xtractor(args.class_number, args.dropout)
    tmp = av_model.state_dict()

    average_param = dict()
    for k in list(asynchronous_model[0].keys()):
        average_param[k] = asynchronous_model[0][k]

        for mod in asynchronous_model[1:]:
            average_param[k] += mod[k]

        if 'num_batches_tracked' not in k:
            tmp[k] = torch.FloatTensor(average_param[k] / len(asynchronous_model))

    # return the file name of the new model
    current_model_file_name = "{}/model_{}_epoch_{}_batch_{}".format(args.model_path, args.expe_id, epoch,
                                                                     megabatch_idx)
    torch.save(tmp, current_model_file_name)
    if megabatch_idx == megabatch_number:
        torch.save(tmp, "{}/model_{}_epoch_{}".format(args.model_path, args.expe_id, epoch))

    return current_model_file_name


Anthony Larcher's avatar
Anthony Larcher committed
279
def train_worker(rank, epoch, args, initial_model_file_name, batch_list, output_queue):
280
281
282
283
284
285
286
287
288
289
290
    """


    :param rank:
    :param epoch:
    :param args:
    :param initial_model_file_name:
    :param batch_list:
    :param output_queue:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
291
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
292
293
294
295
    model.load_state_dict(torch.load(initial_model_file_name))
    model.train()

    torch.manual_seed(args.seed + rank)
Anthony Larcher's avatar
Anthony Larcher committed
296
    train_loader = XvectorMultiDataset(batch_list, args.batch_path)
Anthony Larcher's avatar
Anthony Larcher committed
297
298
299
300
301
302
303
304
305
306
307
308

    device = torch.device("cuda:{}".format(rank))
    model.to(device)

    optimizer = optim.Adam([{'params': model.frame_conv0.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.frame_conv1.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.frame_conv2.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.frame_conv3.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.frame_conv4.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.seg_lin0.parameters(), 'weight_decay': args.l2_seg},
                            {'params': model.seg_lin1.parameters(), 'weight_decay': args.l2_seg},
                            {'params': model.seg_lin2.parameters(), 'weight_decay': args.l2_seg}
Anthony Larcher's avatar
Anthony Larcher committed
309
                            ], lr=args.lr)
Anthony Larcher's avatar
Anthony Larcher committed
310

Anthony Larcher's avatar
Anthony Larcher committed
311
    criterion = torch.nn.CrossEntropyLoss()
Anthony Larcher's avatar
Anthony Larcher committed
312
313
314
315
316
317
318
319

    accuracy = 0.0
    for batch_idx, (data, target) in enumerate(train_loader):
        optimizer.zero_grad()
        output = model(data.to(device))
        loss = criterion(output, target.to(device))
        loss.backward()
        optimizer.step()
Anthony Larcher's avatar
Anthony Larcher committed
320

Anthony Larcher's avatar
Anthony Larcher committed
321
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()
Anthony Larcher's avatar
Anthony Larcher committed
322

Anthony Larcher's avatar
Anthony Larcher committed
323
324
325
326
327
        if batch_idx % args.log_interval == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
                epoch, batch_idx + 1, train_loader.__len__(),
                       100. * batch_idx / train_loader.__len__(), loss.item(),
                       100.0 * accuracy.item() / ((batch_idx + 1) * args.batch_size)))
Anthony Larcher's avatar
Anthony Larcher committed
328

Anthony Larcher's avatar
Anthony Larcher committed
329
330
    model_param = OrderedDict()
    params = model.state_dict()
Anthony Larcher's avatar
Anthony Larcher committed
331

Anthony Larcher's avatar
Anthony Larcher committed
332
333
334
    for k in list(params.keys()):
        model_param[k] = params[k].cpu().detach().numpy()
    output_queue.put(model_param)
Anthony Larcher's avatar
Anthony Larcher committed
335
336


Anthony Larcher's avatar
Anthony Larcher committed
337
def cross_validation(args, current_model_file_name):
Anthony Larcher's avatar
Anthony Larcher committed
338
339
    """

Anthony Larcher's avatar
Anthony Larcher committed
340
341
    :param args:
    :param current_model_file_name:
Anthony Larcher's avatar
Anthony Larcher committed
342
343
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
344
    with open(args.cross_validation_list, 'r') as fh:
Anthony Larcher's avatar
Anthony Larcher committed
345
        cross_validation_list = [l.rstrip() for l in fh]
Anthony Larcher's avatar
Anthony Larcher committed
346
        sub_lists = split_file_list(cross_validation_list, args.num_processes)
Anthony Larcher's avatar
Anthony Larcher committed
347

Anthony Larcher's avatar
Anthony Larcher committed
348
349
    #
    output_queue = mp.Queue()
Anthony Larcher's avatar
Anthony Larcher committed
350

Anthony Larcher's avatar
Anthony Larcher committed
351
352
353
354
355
356
357
358
    processes = []
    for rank in range(args.num_processes):
        p = mp.Process(target=cv_worker,
                       args=(rank, args, current_model_file_name, sub_lists[rank], output_queue)
                       )
        # We first evaluate the model across `num_processes` processes
        p.start()
        processes.append(p)
Anthony Larcher's avatar
Anthony Larcher committed
359

Anthony Larcher's avatar
Anthony Larcher committed
360
361
362
363
    # Average the models and write the new one to disk
    result = []
    for ii in range(args.num_processes):
        result.append(output_queue.get())
Anthony Larcher's avatar
Anthony Larcher committed
364

Anthony Larcher's avatar
Anthony Larcher committed
365
366
    for p in processes:
        p.join()
Anthony Larcher's avatar
Anthony Larcher committed
367

Anthony Larcher's avatar
Anthony Larcher committed
368
369
370
    # Compute the global accuracy
    accuracy = 0.0
    total_batch_number = 0
Anthony Larcher's avatar
Anthony Larcher committed
371
    for bn, acc in result:
Anthony Larcher's avatar
Anthony Larcher committed
372
        accuracy += acc
Anthony Larcher's avatar
Anthony Larcher committed
373
374
        total_batch_number += bn
    
Anthony Larcher's avatar
Anthony Larcher committed
375
    return 100. * accuracy / (total_batch_number * args.batch_size)
Anthony Larcher's avatar
Anthony Larcher committed
376
377


Anthony Larcher's avatar
Anthony Larcher committed
378
def cv_worker(rank, args, current_model_file_name, batch_list, output_queue):
Anthony Larcher's avatar
Anthony Larcher committed
379
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
380
381
    model.load_state_dict(torch.load(current_model_file_name))
    model.eval()
Anthony Larcher's avatar
Anthony Larcher committed
382

Anthony Larcher's avatar
Anthony Larcher committed
383
    cv_loader = XvectorMultiDataset(batch_list, args.batch_path)
Anthony Larcher's avatar
Anthony Larcher committed
384

Anthony Larcher's avatar
Anthony Larcher committed
385
386
    device = torch.device("cuda:{}".format(rank))
    model.to(device)
Anthony Larcher's avatar
Anthony Larcher committed
387

Anthony Larcher's avatar
Anthony Larcher committed
388
    accuracy = 0.0
Anthony Larcher's avatar
Anthony Larcher committed
389
    for batch_idx, (data, target) in enumerate(cv_loader):
Anthony Larcher's avatar
Anthony Larcher committed
390
391
        output = model(data.to(device))
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()
Anthony Larcher's avatar
Anthony Larcher committed
392
    output_queue.put((cv_loader.__len__(), accuracy.cpu().numpy()))
Anthony Larcher's avatar
Anthony Larcher committed
393

Anthony Larcher's avatar
hot    
Anthony Larcher committed
394

395
def extract_idmap(args, device_id, segment_indices, fs_params, idmap_name, output_queue):
Anthony Larcher's avatar
Anthony Larcher committed
396
    """
Anthony Larcher's avatar
Anthony Larcher committed
397
398
    Function that takes a model and an idmap and extract all x-vectors based on this model
    and return a StatServer containing the x-vectors
Anthony Larcher's avatar
Anthony Larcher committed
399
    """
Anthony Larcher's avatar
Anthony Larcher committed
400
    device = torch.device("cuda:{}".format(device_ID))
Anthony Larcher's avatar
Anthony Larcher committed
401
    device = torch.device('cpu')
Anthony Larcher's avatar
Anthony Larcher committed
402
403
404
405
406
407
408
409
410
411
412
413
414

    # Create the dataset
    tmp_idmap = IdMap(idmap_name)
    idmap = IdMap()
    idmap.leftids = tmp_idmap.leftids[segment_indices]
    idmap.rightids = tmp_idmap.rightids[segment_indices]
    idmap.start = tmp_idmap.start[segment_indices]
    idmap.stop = tmp_idmap.stop[segment_indices]

    segment_loader = StatDataset(idmap, fs_params)

    # Load the model
    model_file_name = '/'.join([args.model_path, args.model_name])
Anthony Larcher's avatar
Anthony Larcher committed
415
    model = Xtractor(args.class_number, args.dropout)
Anthony Larcher's avatar
Anthony Larcher committed
416
417
418
419
420
421
422
423
    model.load_state_dict(torch.load(model_file_name))
    model.eval()

    # Get the size of embeddings
    emb_a_size = model.seg_lin0.weight.data.shape[0]
    emb_b_size = model.seg_lin1.weight.data.shape[0]

    # Create a Tensor to store all x-vectors on the GPU
Anthony Larcher's avatar
Anthony Larcher committed
424
425
426
427
428
429
    emb_1 = numpy.zeros((idmap.leftids.shape[0], emb_a_size)).astype(numpy.float32)
    emb_2 = numpy.zeros((idmap.leftids.shape[0], emb_b_size)).astype(numpy.float32)
    emb_3 = numpy.zeros((idmap.leftids.shape[0], emb_b_size)).astype(numpy.float32)
    emb_4 = numpy.zeros((idmap.leftids.shape[0], emb_b_size)).astype(numpy.float32)
    emb_5 = numpy.zeros((idmap.leftids.shape[0], emb_b_size)).astype(numpy.float32)
    emb_6 = numpy.zeros((idmap.leftids.shape[0], emb_b_size)).astype(numpy.float32)
Anthony Larcher's avatar
Anthony Larcher committed
430
431
432
433
434
435

    # Send on selected device
    model.to(device)

    # Loop to extract all x-vectors
    for idx, (model_id, segment_id, data) in enumerate(segment_loader):
Anthony Larcher's avatar
Anthony Larcher committed
436
        logging.critical('Process file {}, [{} / {}]'.format(segment_id, idx, segment_loader.__len__()))
Anthony Larcher's avatar
Anthony Larcher committed
437

Anthony Larcher's avatar
Anthony Larcher committed
438
439
440
        if list(data.shape)[2] < 20:
            pass
        else:
Anthony Larcher's avatar
Anthony Larcher committed
441
442
443
444
445
446
447
            seg_1, seg_2, seg_3, seg_4, seg_5, seg_6 = model.extract(data.to(device))
            emb_1[idx, :] = seg_1.detach().cpu()
            emb_2[idx, :] = seg_2.detach().cpu()
            emb_3[idx, :] = seg_3.detach().cpu()
            emb_4[idx, :] = seg_4.detach().cpu()
            emb_5[idx, :] = seg_5.detach().cpu()
            emb_6[idx, :] = seg_6.detach().cpu()
Anthony Larcher's avatar
Anthony Larcher committed
448

Anthony Larcher's avatar
Anthony Larcher committed
449
    output_queue.put((segment_indices, emb_1, emb_2, emb_3, emb_4, emb_5, emb_6))
Anthony Larcher's avatar
Anthony Larcher committed
450
451


452
453
454
455
def xtrain_single(args):
    """
    Initialize and train an x-vector on a single GPU

Anthony Larcher's avatar
Anthony Larcher committed
456
457
458
    :param args:
    :return:
    """
459
    # If we start from an existing model
Anthony Larcher's avatar
Anthony Larcher committed
460
    if not args.init_model_name == '':
461
        # Load the model
Anthony Larcher's avatar
Anthony Larcher committed
462
463
        logging.critical("*** Load model from = {}/{}".format(args.model_path, args.init_model_name))
        model_file_name = '/'.join([args.model_path, args.init_model_name])
Anthony Larcher's avatar
minor    
Anthony Larcher committed
464
        model = torch.load(model_file_name)
Anthony Larcher's avatar
Anthony Larcher committed
465
        model.train()
466
467
468
469
    else:
        # Initialize a first model and save to disk
        model = Xtractor(args.class_number, args.dropout)
        model.train()
Anthony Larcher's avatar
Anthony Larcher committed
470
    model.cuda()
Anthony Larcher's avatar
Anthony Larcher committed
471

472
473
    # Split the training data in train and cv
    total_seg_df = pickle.load(open(args.batch_training_list, "rb"))
Anthony Larcher's avatar
Anthony Larcher committed
474
475
476
477
478
479

    speaker_dict = {}
    tmp = total_seg_df.speaker_id.unique()
    tmp.sort()
    for idx, spk in enumerate(tmp):
        speaker_dict[spk] = idx
Anthony Larcher's avatar
Anthony Larcher committed
480
481
    pickle.dump(speaker_dict, open( "spk_dictionary.pkl", "wb" ) )

Anthony Larcher's avatar
Anthony Larcher committed
482

Anthony Larcher's avatar
minor    
Anthony Larcher committed
483
    cv_portion = 0.007
484
485
486
487
488
    idx = numpy.arange(len(total_seg_df))
    numpy.random.shuffle(idx)
    train_seg_df = total_seg_df.iloc[idx[:int((1-cv_portion)*len(idx))]].reset_index()
    cv_seg_df = total_seg_df.iloc[idx[int((1-cv_portion)*len(idx)):]].reset_index()

Anthony Larcher's avatar
Anthony Larcher committed
489
490
491
    current_model_file_name = "initial_model"
    torch.save(model.state_dict(), current_model_file_name)

Anthony Larcher's avatar
Anthony Larcher committed
492
493
    initial_lr = args.lr

Anthony Larcher's avatar
Anthony Larcher committed
494
495
    for epoch in range(1, args.epochs + 1):
        # Process one epoch and return the current model
Anthony Larcher's avatar
Anthony Larcher committed
496
        model = train_epoch_single(model, epoch, train_seg_df, speaker_dict, args)
Anthony Larcher's avatar
Anthony Larcher committed
497
498

        # Add the cross validation here
Anthony Larcher's avatar
Anthony Larcher committed
499
        accuracy = cross_validation_single(args, model, cv_seg_df, speaker_dict)
Anthony Larcher's avatar
Anthony Larcher committed
500
501
502
503
        logging.critical("*** Cross validation accuracy = {} %".format(accuracy))

        # Decrease learning rate after every epoch
        args.lr = args.lr * 0.9
Anthony Larcher's avatar
Anthony Larcher committed
504
505
        if epoch == 50:
            args.lr = initial_lr
Anthony Larcher's avatar
Anthony Larcher committed
506
507
508
        logging.critical("        Decrease learning rate: {}".format(args.lr))

        # return the file name of the new model
Anthony Larcher's avatar
Anthony Larcher committed
509
510
511
512
        base_name = "model"
        if not args.init_model_name == "":
            base_name = args.init_model_name
        current_model_file_name = "{}/{}_{}_epoch_{}".format(base_name, args.model_path, args.expe_id, epoch)
Anthony Larcher's avatar
Anthony Larcher committed
513
514
515
        torch.save(model, current_model_file_name)


Anthony Larcher's avatar
Anthony Larcher committed
516
def train_epoch_single(model, epoch, train_seg_df, speaker_dict, args):
Anthony Larcher's avatar
Anthony Larcher committed
517
518
519
520
    """

    :param model:
    :param epoch:
521
    :param train_seg_df:
Anthony Larcher's avatar
Anthony Larcher committed
522
523
524
    :param args:
    :return:
    """
525
    device = torch.device("cuda:0")
Anthony Larcher's avatar
Anthony Larcher committed
526
    
Anthony Larcher's avatar
Anthony Larcher committed
527
    torch.manual_seed(args.seed)
Anthony Larcher's avatar
Anthony Larcher committed
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542

    train_transform = []
    if not args.train_transformation == '':
        trans = args.train_transformation.split(',')
        for t in trans:
            if "CMVN" in t:
                train_transform.append(CMVN())
            if "FrequencyMask" in t:
                a = t.split(",")[0].split("(")[1]
                b = t.split(",")[1].split("(")[0]
                train_transform.append(FrequencyMask(a, b))
            if "TemporalMask" in t:
                a = t.split(",")[0].split("(")[1]
                train_transform.append(TemporalMask(a, b))
    train_set = VoxDataset(train_seg_df, speaker_dict, 500, transform=transforms.Compose(train_transform),
Anthony Larcher's avatar
Anthony Larcher committed
543
                           spec_aug_ratio=args.spec_aug, temp_aug_ratio=args.temp_aug)
Anthony Larcher's avatar
minor    
Anthony Larcher committed
544
    train_loader = DataLoader(train_set, batch_size=args.batch_size, shuffle=True, num_workers=15)
Anthony Larcher's avatar
Anthony Larcher committed
545
546
547
548
549
550
551
552
553
554
555
556
557

    optimizer = optim.Adam([{'params': model.frame_conv0.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.frame_conv1.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.frame_conv2.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.frame_conv3.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.frame_conv4.parameters(), 'weight_decay': args.l2_frame},
                            {'params': model.seg_lin0.parameters(), 'weight_decay': args.l2_seg},
                            {'params': model.seg_lin1.parameters(), 'weight_decay': args.l2_seg},
                            {'params': model.seg_lin2.parameters(), 'weight_decay': args.l2_seg}
                            ], lr=args.lr)
    criterion = torch.nn.CrossEntropyLoss()

    accuracy = 0.0
Anthony Larcher's avatar
SpecAug    
Anthony Larcher committed
558
    for batch_idx, (data, target, _, __) in enumerate(train_loader):
Anthony Larcher's avatar
Anthony Larcher committed
559
        target = target.squeeze()
Anthony Larcher's avatar
Anthony Larcher committed
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
        optimizer.zero_grad()
        output = model(data.to(device))
        loss = criterion(output, target.to(device))
        loss.backward()
        optimizer.step()
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()

        if batch_idx % args.log_interval == 0:
            logging.critical('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.3f}'.format(
                             epoch, batch_idx + 1, train_loader.__len__(),
                             100. * batch_idx / train_loader.__len__(), loss.item(),
                             100.0 * accuracy.item() / ((batch_idx + 1) * args.batch_size)))
    return model


Anthony Larcher's avatar
Anthony Larcher committed
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
def cross_validation(args, model):

    with open(args.cross_validation_list, 'r') as fh:
        cross_validation_list = [l.rstrip() for l in fh]
    cv_loader = XvectorMultiDataset(cross_validation_list, args.batch_path)

    model.eval()
    device = torch.device("cuda:0")
    model.to(device)

    accuracy = 0.0
    for batch_idx, (data, target) in enumerate(cv_loader):
        output = model(data.to(device))
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()
    return 100. * accuracy.cpu().numpy() / ((batch_idx + 1) * args.batch_size)


Anthony Larcher's avatar
Anthony Larcher committed
592
def cross_validation_single(args, model, cv_seg_df, speaker_dict):
593
594
595
596
597
598
599
    """

    :param args:
    :param model:
    :param cv_seg_df:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
    cv_transform = []
    if not args.cv_transformation == '':
        trans = args.cv_transformation.split(',')
        for t in trans:
            if "CMVN" in t:
                cv_transform.append(CMVN())
            if "FrequencyMask" in t:
                a = t.split(",")[0].split("(")[1]
                b = t.split(",")[1].split("(")[0]
                cv_transform.append(FrequencyMask(a, b))
            if "TemporalMask" in t:
                a = t.split(",")[0].split("(")[1]
                cv_transform.append(TemporalMask(a, b))
    cv_set = VoxDataset(cv_seg_df, speaker_dict, 500, transform=transforms.Compose(cv_transform),
                        spec_aug_ratio=args.spec_aug, temp_aug_ratio=args.temp_aug)
Anthony Larcher's avatar
minor    
Anthony Larcher committed
615
    cv_loader = DataLoader(cv_set, batch_size=args.batch_size, shuffle=False, num_workers=15)
616
617
618
619
620
    model.eval()
    device = torch.device("cuda:0")
    model.to(device)

    accuracy = 0.0
Anthony Larcher's avatar
minor    
Anthony Larcher committed
621
    print(cv_set.__len__())
Anthony Larcher's avatar
SpecAug    
Anthony Larcher committed
622
    for batch_idx, (data, target, _, __) in enumerate(cv_loader):
Anthony Larcher's avatar
Anthony Larcher committed
623
        target = target.squeeze()
624
625
626
627
628
        output = model(data.to(device))
        accuracy += (torch.argmax(output.data, 1) == target.to(device)).sum()
    return 100. * accuracy.cpu().numpy() / ((batch_idx + 1) * args.batch_size)


Anthony Larcher's avatar
Anthony Larcher committed
629
def extract_parallel(args, fs_params):
630
631
632
633
634
635
    """

    :param args:
    :param fs_params:
    :return:
    """
Anthony Larcher's avatar
Anthony Larcher committed
636
637
638
    emb_a_size = 512
    emb_b_size = 512

Anthony Larcher's avatar
Anthony Larcher committed
639
    idmap = IdMap(args.idmap)
Anthony Larcher's avatar
Anthony Larcher committed
640

Anthony Larcher's avatar
Anthony Larcher committed
641
642
643
644
645
646
647
648
649
650
651
652
653
    x_server_1 = StatServer(idmap, 1, emb_a_size)
    x_server_2 = StatServer(idmap, 1, emb_b_size)
    x_server_3 = StatServer(idmap, 1, emb_b_size)
    x_server_4 = StatServer(idmap, 1, emb_b_size)
    x_server_5 = StatServer(idmap, 1, emb_b_size)
    x_server_6 = StatServer(idmap, 1, emb_b_size)

    x_server_1.stat0 = numpy.ones(x_server_1.stat0.shape)
    x_server_2.stat0 = numpy.ones(x_server_2.stat0.shape)
    x_server_3.stat0 = numpy.ones(x_server_3.stat0.shape)
    x_server_4.stat0 = numpy.ones(x_server_4.stat0.shape)
    x_server_5.stat0 = numpy.ones(x_server_5.stat0.shape)
    x_server_6.stat0 = numpy.ones(x_server_6.stat0.shape)
Anthony Larcher's avatar
Anthony Larcher committed
654
655
656

    # Split the indices
    mega_batch_size = idmap.leftids.shape[0] // args.num_processes
Anthony Larcher's avatar
Anthony Larcher committed
657
658
659

    logging.critical("Number of sessions to process: {}".format(idmap.leftids.shape[0]))

Anthony Larcher's avatar
Anthony Larcher committed
660
661
662
    segment_idx = []
    for ii in range(args.num_processes):
        segment_idx.append(
Anthony Larcher's avatar
Anthony Larcher committed
663
664
665
666
            numpy.arange(ii * mega_batch_size, numpy.min([(ii + 1) * mega_batch_size, idmap.leftids.shape[0]])))

    for idx, si in enumerate(segment_idx):
        logging.critical("Number of session on process {}: {}".format(idx, len(si)))
Anthony Larcher's avatar
Anthony Larcher committed
667
668
669
670
671
672
673

    # Extract x-vectors in parallel
    output_queue = mp.Queue()

    processes = []
    for rank in range(args.num_processes):
        p = mp.Process(target=extract_idmap,
Anthony Larcher's avatar
Anthony Larcher committed
674
                       args=(args, rank, segment_idx[rank], fs_params, args.idmap, output_queue)
Anthony Larcher's avatar
Anthony Larcher committed
675
676
677
678
679
680
681
                       )
        # We first train the model across `num_processes` processes
        p.start()
        processes.append(p)

    # Get the x-vectors and fill the StatServer
    for ii in range(args.num_processes):
Anthony Larcher's avatar
Anthony Larcher committed
682
683
684
685
686
687
688
        indices, seg_1, seg_2, seg_3, seg_4, seg_5, seg_6 = output_queue.get()
        x_server_1.stat1[indices, :] = seg_1
        x_server_2.stat1[indices, :] = seg_2
        x_server_3.stat1[indices, :] = seg_3
        x_server_4.stat1[indices, :] = seg_4
        x_server_5.stat1[indices, :] = seg_5
        x_server_6.stat1[indices, :] = seg_6
Anthony Larcher's avatar
Anthony Larcher committed
689
690
691
692

    for p in processes:
        p.join()

Anthony Larcher's avatar
Anthony Larcher committed
693
    return x_server_1, x_server_2, x_server_3, x_server_4, x_server_5, x_server_6
Anthony Larcher's avatar
Anthony Larcher committed
694
695
696
697
698