/
main_visual_lip_lrs3_new.py
456 lines (398 loc) · 22.5 KB
/
main_visual_lip_lrs3_new.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
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
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
import wandb, shutil
import numpy as np
import torch
import torch.optim as optim
import torch.nn.functional as F
import os, yaml, pprint, warnings
from yaml import Loader as CLoader
from tqdm import tqdm
import torch.multiprocessing as mp
from multiprocessing import Process, Manager
from torch.utils.data import DataLoader
import datasets_new as datasets
from models.loss import AAMsoftmax
from models.network_new import AudioModel, VisualModel, AudioVisualModel
from utils.eval_metrics_new import *
################# GLOBAL CONFIGURATION ##############################
MODE = 'test' # train | test | finetune
#####################################################################
SEED = 1022 # fixed random seed for fair comparison
np.random.seed(SEED)
torch.manual_seed(SEED)
torch.cuda.manual_seed_all(SEED)
warnings.filterwarnings('ignore')
with open('./conf/savedsolution_visual_lrs3_IT.yaml', 'r') as f:
OPTS = yaml.load(f, Loader=CLoader)
# shutil.copyfile('./conf/config_visual_lip_lrs3_new.yaml', './conf/savedsolution_visual_lrs3_IT.yaml')
# Basic functions for train and test phase
def get_visualmodel(model):
return VisualModel(model)
class Trainer(object):
def __init__(self):
self.stageopts = OPTS[MODE]
self.visualmodelname = self.stageopts['visualmodel']
self.visualmodelopts = OPTS[self.visualmodelname]
self.visualresume = self.stageopts['visualresume']
self.visualemb_dim = self.visualmodelopts['embedding_dim']
# neural network
visualmodel = get_visualmodel(self.visualmodelname)
print('visual model parameters_count: %.2fM' % (sum(p.numel() for p in visualmodel.parameters() if p.requires_grad)/1e6))
device_ids, device_num = self.stageopts['gpus'], len(self.stageopts['gpus'])
os.environ['CUDA_VISIBLE_DEVICES'] = ','.join([str(gpu) for gpu in device_ids])
self.device_ids = [x for x in range(len(device_ids))]
self.device = torch.device('cuda') # : +str(device_ids[0])
self.visualmodel = torch.nn.DataParallel(visualmodel.to(self.device), device_ids=self.device_ids)
self.dataopts = {}
# self.dataopts = {**{'seconds':self.stageopts['seconds']}}
self.dataopts = {'seconds':self.stageopts['seconds']}
for traindata in ['train_manifest', 'train_videodir']: # for train, val, and aug
# self.dataopts = {**self.dataopts, **{traindata:OPTS[traindata]}}
self.dataopts[traindata] = OPTS[traindata]
for valdata in ['test_trial', 'test_videodir']: # for val
# self.dataopts = {**self.dataopts, **{valdata:OPTS['test_lrs3'][valdata]}}
self.dataopts[valdata] = OPTS['test_lrs3'][valdata]
self.trainset = datasets.VisualLipTrainset(self.dataopts)
self.trainloader = DataLoader(self.trainset, shuffle=True, batch_size=self.stageopts['batchsize'], num_workers=4*device_num, drop_last=True)
self.visualcriterion = AAMsoftmax(n_class=self.trainset.n_spk, m=self.stageopts['margins'][1], s=self.stageopts['scale'], em_dim=self.visualemb_dim).to(self.device)
param_groups = [{'params': self.visualmodel.parameters()},
{'params': self.visualcriterion.parameters()},
]
if self.stageopts['optimizer'] == 'sgd':
self.optimopts = OPTS['sgd']
self.optim = optim.SGD(param_groups, self.optimopts['init_lr'], nesterov=self.optimopts['nesterov'], momentum = self.optimopts['momentum'], weight_decay = self.optimopts['weight_decay'])
elif self.stageopts['optimizer'] == 'adam':
self.optimopts = OPTS['adam']
self.optim = optim.Adam(param_groups, lr=self.optimopts['init_lr'] , weight_decay=self.optimopts['weight_decay'])
if self.stageopts['lr_scheduler'] == 'steplr':
# self.lr_scheduler = optim.lr_scheduler.StepLR(self.optim, step_size=4, gamma=0.5)
# self.lr_scheduler = optim.lr_scheduler.StepLR(self.optim, step_size=1, gamma=0.97)
self.lr_scheduler = optim.lr_scheduler.MultiStepLR(self.optim, milestones=[10,15], gamma = 0.1)
elif self.stageopts['lr_scheduler'] == 'cycliclr':
self.lr_scheduler = optim.lr_scheduler.CyclicLR(self.optim,cycle_momentum=False,base_lr=0.000001,max_lr=0.001,step_size_up=2000,step_size_down=2000)
self.eers = []
self.dcfs = []
self.current_epoch, self.epochs = 0, self.stageopts['epochs']
if self.visualresume == 'None':
self.exp = '_'.join(['lrs3', 'independent-training', self.visualmodelname, 'noisespec3blocks1'])
else:
self.exp = self.visualresume.split('/')[1]
# continue training
if (self.visualresume != 'None' and os.path.exists(self.visualresume)):
self._load()
else:
print('Train from scratch: there is no specified audio resume.')
self.wandb = self.stageopts['wandb']
if self.wandb != False:
config = {
"learning_rate": self.optimopts['init_lr'],
"epochs": self.epochs,
"batch_size": self.stageopts['batchsize']
}
wandb.init(project="Short-Short-New", config=config)
def _plan(self):
valuerate = max(((self.current_epoch - 1)/(20 - 1)) * (0.8 - 1) + 1, 0.8)
return valuerate
def _load(self):
print('loading visual model from {}'.format(self.visualresume))
if not os.path.exists(self.visualresume):
print('No pretrained visual model exists!')
else:
visualckpt = torch.load(self.visualresume)
if 'visualstate_dict' in visualckpt.keys():
self.visualmodel.load_state_dict(visualckpt['visualstate_dict'], strict=False)
if 'visualcriterion' in visualckpt.keys():
self.visualcriterion = visualckpt['visualcriterion']
def _train(self):
start_epoch = self.current_epoch
for epoch in range(start_epoch + 1, self.epochs + 1):
self.current_epoch = epoch
self._train_epoch()
self.lr_scheduler.step()
def _train_epoch(self):
self.visualmodel.train()
self.visualcriterion.train()
videosum_loss, videosum_samples, videocorrect = 0, 0, 0
progress_bar = tqdm(self.trainloader)
for batch_idx, (videofeats, targets_label) in enumerate(progress_bar):
self.optim.zero_grad()
videofeats = videofeats.to(self.device)
targets_label = targets_label.to(self.device)
frame_video, emb_video = self.visualmodel(videofeats)
vloss, videologits = self.visualcriterion(emb_video, targets_label)
vloss = torch.mean(vloss)
vloss.backward()
videosum_samples += len(videofeats)
_, videoprediction = torch.max(videologits, dim=1)
videocorrect += (videoprediction == targets_label).sum().item()
# if self.current_epoch == 1 and batch_idx < 2000:
# lr_scale = min(1., float(batch_idx + 1) / float(2000))
# for _, pg in enumerate(self.optim.param_groups):
# pg['lr'] = lr_scale * 0.001
self.optim.step()
videosum_loss += vloss.item() * len(targets_label)
progress_bar.set_description(
'Train Epoch: {:3d} [{:4d}/{:4d} ({:3.3f}%)] videoLoss: {:.4f} videoAcc: {:.4f}%' #
.format(self.current_epoch, batch_idx + 1,
len(self.trainloader), 100. * (batch_idx + 1) / len(self.trainloader),
videosum_loss / videosum_samples, 100. * videocorrect / videosum_samples))
if self.wandb != False:
wandb.log({"videoLoss":videosum_loss/videosum_samples,
"lr": self.optim.state_dict()['param_groups'][0]['lr'],
"videoAcc":100.*videocorrect/videosum_samples,
})
self._save('exp/{}/net_{}.pth'.format(self.exp, self.current_epoch))
# flexible log
interval_val = 1
if self.current_epoch % interval_val == 0:
eer, minDCF = self._eval_network(stage='val', num_workers=2)
self.eers.append(eer)
self.dcfs.append(minDCF)
if not os.path.isdir('log/'):
os.mkdir('log')
with open('log/'+self.exp+'-training.log', "a+") as score_file:
score_file.write("%d epoch, LR %f, LOSS %f, ACC %2.2f%%, EER %2.2f%%, minDCF %2.4f, bestEER %2.2f%%, bestminDCF %2.4f\n" \
%(self.current_epoch, self.optim.state_dict()['param_groups'][0]['lr'], videosum_loss/videosum_samples, \
100.*videocorrect/videosum_samples, self.eers[-1], self.dcfs[-1], min(self.eers), min(self.dcfs)))
score_file.flush()
def _save(self, modelpath):
torch.save({'visualstate_dict': self.visualmodel.state_dict(),
'visualcriterion': self.visualcriterion},
modelpath.replace('net_','visualnet_'))
def _extract_embedding(self, stage, filelist): # stage='val'
testset = datasets.VisualLipTestset(self.dataopts, filelist, stage='val')
testloader = DataLoader(testset, batch_size=1, shuffle=False, num_workers=1)
localmodel = self.visualmodel.module
localmodel = localmodel.to(self.device)
localmodel.eval()
# print('Extracting test embeddings for {}: '.format(dataset))
emb_dir = os.path.join('exp/{}/{}_emb'.format(self.exp, stage))
os.makedirs(emb_dir, exist_ok = True)
with torch.no_grad():
for signal_global, signal_local, utt in tqdm(testloader):
utt = utt[0]
spk_dir = os.path.join(emb_dir, os.path.dirname(utt))
os.makedirs(spk_dir, exist_ok=True)
signal_global = signal_global.to(self.device)
signal_local = signal_local.squeeze(0).to(self.device)
_, embedding_global = localmodel(signal_global)
_, embedding_local = localmodel(signal_local)
embedding_global = embedding_global.cpu().numpy()
embedding_local = embedding_local.cpu().numpy()
np.savez_compressed(os.path.join(spk_dir, os.path.basename(utt.split('.')[0])), [embedding_global, embedding_local])
del localmodel
torch.cuda.empty_cache()
def _score_embedding(self, stage, trials, score_dict):
# GRID scoring for global and local embeddings and save score in score/MODELDIR/score_TRIAL.txt
emb_dir = os.path.join('exp/{}/{}_emb'.format(self.exp, stage))
for line in trials:
embedding_11 = torch.FloatTensor(np.load(os.path.join(emb_dir, line.split()[-2].split('.')[0]+'.npz'), allow_pickle=True)['arr_0'][0])
embedding_12 = torch.FloatTensor(np.load(os.path.join(emb_dir, line.split()[-2].split('.')[0]+'.npz'), allow_pickle=True)['arr_0'][1])
embedding_21 = torch.FloatTensor(np.load(os.path.join(emb_dir, line.split()[-1].split('.')[0]+'.npz'), allow_pickle=True)['arr_0'][0])
embedding_22 = torch.FloatTensor(np.load(os.path.join(emb_dir, line.split()[-1].split('.')[0]+'.npz'), allow_pickle=True)['arr_0'][1])
embedding_11, embedding_12, embedding_21, embedding_22 = F.normalize(embedding_11),F.normalize(embedding_12),F.normalize(embedding_21),F.normalize(embedding_22)
# Compute the scores
score_1 = torch.mean(torch.matmul(embedding_11, embedding_21.T)) # higher is positive
score_2 = torch.mean(torch.matmul(embedding_12, embedding_22.T))
score = (score_1 + score_2) / 2
score = score.detach().cpu().numpy()
score_dict[line] = [score, int(line.split()[0])] # score, label
def _eval_network(self, stage, num_workers=1):
# prepare for multiprocessing of embedding extraction
files, utts = [], []
trial = self.dataopts['test_trial']
lines = open(trial).read().splitlines()
for line in lines:
files.append(line.split()[1])
files.append(line.split()[2])
setfiles = list(set(files))
setfiles.sort()
part = list(range(0, len(setfiles)+1, int(len(setfiles)//num_workers)))
part[-1] = len(setfiles)
utts = [setfiles[part[i]:part[i+1]] for i in range(num_workers)]
self.visualmodel.cpu()
torch.cuda.empty_cache()
if num_workers == 1:
self._extract_embedding(stage, utts[0])
else:
# extract embeddings and save in exp/MODELDIR/test_emb/
args = [(stage, utts[i]) for i in range(num_workers)]
ctx = mp.get_context("spawn")
jobs = [ctx.Process(target=self._extract_embedding, args=(a)) for a in args]
for j in jobs: j.start()
for j in jobs: j.join()
self.visualmodel = self.visualmodel.to(self.device)
num_workers = 20
score_dict = Manager().dict()
part = list(range(0, len(lines)+1, int(len(lines)//num_workers)))
part[-1] = len(lines)
trials = [lines[part[i]:part[i+1]] for i in range(num_workers)]
args = [(stage, trials[i], score_dict) for i in range(num_workers)]
jobs = [Process(target=self._score_embedding, args=(a)) for a in args]
for j in jobs: j.start()
for j in jobs: j.join()
scores, labels = [], []
for line in lines:
scores.append(score_dict[line][0])
labels.append(score_dict[line][1])
# Coumpute EER and minDCF
EER = tuneThresholdfromScore(scores, labels, [1, 0.1])[1]
fnrs, fprs, thresholds = ComputeErrorRates(scores, labels)
minDCF, _ = ComputeMinDcf(fnrs, fprs, thresholds, 0.05, 1, 1)
print("EER: {:.6f}%, minDCF: {:.6f}".format(EER, minDCF))
return EER, minDCF
def _print_config(self, opts):
pp = pprint.PrettyPrinter(indent = 2)
pp.pprint(opts)
def __call__(self):
print("[Model is saved in: {}]".format(self.exp))
# print("Data opts: ")
# self._print_config(self.dataopts)
# print("Model opts: ")
# self._print_config(self.modelopts)
# print("Stage opts: ")
# self._print_config(self.stageopts)
os.makedirs('exp/{}'.format(self.exp), exist_ok = True)
self._train()
class Tester(object):
def __init__(self):
self.stageopts = OPTS[MODE]
self.visualmodelname = self.stageopts['visualmodel']
self.visualmodelopts = OPTS[self.visualmodelname]
self.visualemb_dim = self.visualmodelopts['embedding_dim']
self.visualresume = self.stageopts['visualresume']
self.exp = self.visualresume.split('/')[1]
self.gpus = self.stageopts['gpus']
self.device_ids = [x for x in range(len(self.gpus))]
self.dataopts = {}
for data in ['train_manifest', 'cohort_manifest']: # for submean, cohort, and test
# self.dataopts = {**self.dataopts, **{data:OPTS[data]}}
self.dataopts[data] = OPTS[data]
for testdata in ['test_trial', 'test_videodir',]: # for test
self.dataopts[testdata] = OPTS[self.stageopts['data']][testdata]
def _extract_embedding(self, stage, filelist, gpu): # test | cohort | submean
testset = datasets.VisualLipTestset(self.dataopts, filelist, stage)
testloader = DataLoader(testset, batch_size=1, shuffle=False, num_workers=1)
os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu)
device = torch.device('cuda')
testmodel = get_visualmodel(self.visualmodelname)
testmodel = torch.nn.DataParallel(testmodel.to(device), device_ids = [0])
visualckpt = torch.load(self.visualresume)
if 'visualstate_dict' in visualckpt.keys():
testmodel.load_state_dict(visualckpt['visualstate_dict'], strict=False)
testmodel.module.eval()
# print('Extracting test embeddings for {}: '.format(dataset))
emb_dir = os.path.join('exp/{}/{}_emb'.format(self.exp, stage))
os.makedirs(emb_dir, exist_ok = True)
with torch.no_grad():
for signal_global, signal_local, utt in tqdm(testloader):
utt = utt[0]
spk_dir = os.path.join(emb_dir, os.path.dirname(utt))
os.makedirs(spk_dir, exist_ok=True)
signal_global = signal_global.to(device)
signal_local = signal_local.squeeze(0).to(device)
_, embedding_global = testmodel(signal_global)
_, embedding_local = testmodel(signal_local)
embedding_global = embedding_global.cpu().numpy()
embedding_local = embedding_local.cpu().numpy()
np.savez_compressed(os.path.join(spk_dir, os.path.basename(utt.split('.')[0])), [embedding_global, embedding_local])
def _score_embedding(self, stage, trials, score_dict):
emb_dir = os.path.join('exp/{}/{}_emb'.format(self.exp, stage))
for line in trials:
embedding_11 = torch.FloatTensor(np.load(os.path.join(emb_dir, line.split()[-2].split('.')[0]+'.npz'), allow_pickle=True)['arr_0'][0])
embedding_12 = torch.FloatTensor(np.load(os.path.join(emb_dir, line.split()[-2].split('.')[0]+'.npz'), allow_pickle=True)['arr_0'][1])
embedding_21 = torch.FloatTensor(np.load(os.path.join(emb_dir, line.split()[-1].split('.')[0]+'.npz'), allow_pickle=True)['arr_0'][0])
embedding_22 = torch.FloatTensor(np.load(os.path.join(emb_dir, line.split()[-1].split('.')[0]+'.npz'), allow_pickle=True)['arr_0'][1])
embedding_11, embedding_12, embedding_21, embedding_22 = F.normalize(embedding_11),F.normalize(embedding_12),F.normalize(embedding_21),F.normalize(embedding_22)
# Compute the scores
score_1 = torch.mean(torch.matmul(embedding_11, embedding_21.T)) # higher is positive
score_2 = torch.mean(torch.matmul(embedding_12, embedding_22.T))
score = (score_1 + score_2) / 2
score = score.detach().cpu().numpy()
if len(line.split()) == 3:
score_dict[line] = [score, int(line.split()[0])]
elif len(line.split()) == 2:
score_dict[line] = [score]
def _eval_network(self, stage, num_workers=1):
# prepare for multiprocessing of embedding extraction
files, utts = [], []
if stage == 'test':
trial = self.dataopts['test_trial']
lines = open(trial).read().splitlines()
for line in lines:
files.append(line.split()[-2])
files.append(line.split()[-1])
setfiles = list(set(files))
setfiles.sort()
part = list(range(0, len(setfiles)+1, int(len(setfiles)//num_workers)))
part[-1] = len(setfiles)
utts = [setfiles[part[i]:part[i+1]] for i in range(num_workers)]
elif stage == 'cohort':
trial = self.dataopts['cohort_manifest']
lines = open(trial).read().splitlines()
for line in lines:
files.append(line)
setfiles = list(set(files))
setfiles.sort()
part = list(range(0, len(setfiles)+1, int(len(setfiles)//num_workers)))
part[-1] = len(setfiles)
utts = [setfiles[part[i]:part[i+1]] for i in range(num_workers)]
print('loading visual model from {}'.format(self.visualresume))
if num_workers == 1:
self._extract_embedding(stage, utts[0], 0)
else:
# extract embeddings and save in exp/MODELDIR/test_emb/
args = [(stage, utts[i], self.gpus[i%len(self.gpus)]) for i in range(num_workers)]
jobs = [Process(target=self._extract_embedding, args=(a)) for a in args]
for j in jobs: j.start()
for j in jobs: j.join()
if stage == 'cohort':
print('finished')
return
num_workers = 20
score_dict = Manager().dict()
part = list(range(0, len(lines)+1, int(len(lines)//num_workers)))
part[-1] = len(lines)
trials = [lines[part[i]:part[i+1]] for i in range(num_workers)]
args = [(stage, trials[i], score_dict) for i in range(num_workers)]
jobs = [Process(target=self._score_embedding, args=(a)) for a in args]
for j in jobs: j.start()
for j in jobs: j.join()
# GRID scoring for global and local embeddings and save score in score/MODELDIR/score_TRIAL.txt
scores, labels = [], []
score_fname = os.path.join('score', self.exp, 'score_'+os.path.basename(trial))
os.makedirs(os.path.join('score', self.exp), exist_ok = True)
with open(score_fname, 'w+') as fout:
for line in lines:
scores.append(score_dict[line][0])
if len(line.split()) == 3:
labels.append(score_dict[line][1])
if self.stageopts['write_score']:
if len(line.split()) == 3:
fout.write('{} {} {:.5f} {}\n'.format(line.split()[-2], line.split()[-1], score_dict[line][0], line.split()[0]))
elif len(line.split()) == 2:
fout.write('{} {} {:.5f}\n'.format(line.split()[-2], line.split()[-1], score_dict[line][0]))
if len(line.split()) == 3:
# Coumpute EER and minDCF
EER = tuneThresholdfromScore(scores, labels, [1, 0.1])[1]
fnrs, fprs, thresholds = ComputeErrorRates(scores, labels)
minDCF, _ = ComputeMinDcf(fnrs, fprs, thresholds, 0.05, 1, 1)
print("EER: {:.6f}%, minDCF: {:.6f}".format(EER, minDCF))
def _print_config(self, opts):
pp = pprint.PrettyPrinter(indent = 2)
pp.pprint(opts)
if __name__ == '__main__':
if MODE == 'train' or MODE == 'finetune':
trainer = Trainer()
trainer()
# voxceleb dataset test
elif MODE == 'test':
tester = Tester()
tester._eval_network(stage='test', num_workers=4)
# elif MODE == 'extract_submean_embedding':
# trainer.extract_submeanembedding()
# elif MODE == 'featurefusion':
# eer, minDCF = utils.eer_cos_featurefusion(trainer.log_time, dataset='lrs3', trial=trainer.data_opts['test_lrs3']['test_trial_lrs3'])
# print("EER: {:.6f}%, minDCF: {:.6f}".format(eer * 100, minDCF))
# elif MODE == 'scorefusion':
# eer, minDCF = utils.eer_cos_scorefusion('score_v_FaceResNet101-JointTraining.txt','score_v_MiniFace-JointTraining.txt')
# print("EER: {:.6f}%, minDCF: {:.6f}".format(eer * 100, minDCF))