-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathasv.py
169 lines (136 loc) · 8.07 KB
/
asv.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
# This code is partly based on
# https://github.com/speechbrain/speechbrain/blob/develop/recipes/VoxCeleb/SpeakerRec/speaker_verification_plda.py
import logging
from pathlib import Path
import torch
from speechbrain.utils.metric_stats import EER
from sklearn.metrics.pairwise import cosine_distances
import pandas as pd
from anonymization.modules.speaker_embeddings.anonymization.utils.plda_model import PLDAModel
from anonymization.modules.speaker_embeddings import SpeakerExtraction
from utils import write_table, read_kaldi_format, save_kaldi_format
logger = logging.getLogger(__name__)
class ASV:
def __init__(self, model_dir, device, score_save_dir, distance='plda', plda_settings=None, vec_type='xvector'):
self.device = device
self.vec_type = vec_type
self.model_dir = model_dir
self.score_save_dir = score_save_dir
self.distance = distance
if plda_settings:
self.plda_model_dir = plda_settings['model_dir']
self.plda_train_data_dir = plda_settings['train_data_dir']
self.plda_anon = plda_settings['anon'] # whether this model is trained on anon data or original
else:
if self.distance == 'plda':
raise KeyError('PLDA settings must be given in config when using distance=plda!')
self.plda_model_dir = None
self.plda_train_data_dir = None
self.plda_anon = None
self.extractor = SpeakerExtraction(results_dir=self.score_save_dir / 'emb_xvect',
devices=[self.device],
settings={'vec_type': vec_type, 'emb_level': 'utt', 'emb_model_path': model_dir})
def compute_trial_scores(self, trials, enrol_indices, test_indices, out_file, sim_scores):
scores = []
for enrol_id, test_id in trials:
enrol_index = enrol_indices[enrol_id]
test_index = test_indices[test_id]
s = float(sim_scores[enrol_index, test_index])
scores.append([enrol_id, test_id, s])
score_data = pd.DataFrame(scores, columns=['enroll_id', 'trial_id', 'score'])
write_table(filename=out_file, table=score_data, sep=' ')
return score_data
def _split_scores(self, scores, labels):
positive_scores = []
negative_scores = []
for enrol_id, test_id, score in scores:
label = labels[(enrol_id, test_id)]
if label == 1:
positive_scores.append(score)
else:
negative_scores.append(score)
return positive_scores, negative_scores
def select_data_for_plda(self, all_data_dir, selected_data_dir, out_dir):
def change_id_format(data_dict):
return {k.replace('--', '-'): v for k, v in data_dict.items()}
df = pd.read_csv(selected_data_dir / 'train.csv', sep=',')
selected_utts = set([change_id_format(segment.split('_')[0]) for segment in df['ID'].to_list()])
wav_scp = read_kaldi_format(all_data_dir / 'wav.scp')
utt2spk = change_id_format(read_kaldi_format(all_data_dir / 'utt2spk'))
spk2gender = change_id_format(read_kaldi_format(all_data_dir / 'spk2gender'))
selected_wav_scp = {utt: wav for utt, wav in wav_scp.items() if utt in selected_utts}
selected_utt2spk = {utt: spk for utt, spk in utt2spk.items() if utt in selected_utts}
selected_spk2gender = {spk: spk2gender[spk] for spk in set(selected_utt2spk.values())}
out_dir.mkdir(parents=True, exist_ok=True)
save_kaldi_format(selected_wav_scp, out_dir / 'wav.scp')
save_kaldi_format(selected_utt2spk, out_dir / 'utt2spk')
save_kaldi_format(selected_spk2gender, out_dir / 'spk2gender')
def eer_compute(self, enrol_dir, test_dir, trial_runs_file):
# Compute all enrol(spk level) and Test(utt level) embeddings
# enroll vectors are the speaker-level average vectors
enrol_all_dict = self.extractor.extract_speakers(dataset_path=Path(enrol_dir), emb_level='spk')
test_all_dict = self.extractor.extract_speakers(dataset_path=Path(test_dir), emb_level='utt')
enrol_vectors = []
enrol_ids = []
test_vectors = []
test_ids = []
trials = {}
# 1462 1462-170142-0000 target
# 1462 2412-153948-0004 nontarget
with open(trial_runs_file, 'r') as f:
for line in f:
temp = line.strip().split(' ')
if temp[0] not in set(enrol_ids):
enrol_vectors.append(enrol_all_dict.get_embedding_for_identifier(temp[0]))
enrol_ids.append(temp[0])
if temp[1] not in set(test_ids):
test_vectors.append(test_all_dict.get_embedding_for_identifier(temp[1]))
test_ids.append(temp[1])
trials[(temp[0], temp[1])] = int(temp[2] == 'target')
enrol_vectors = torch.stack(enrol_vectors)
test_vectors = torch.stack(test_vectors)
save_dir = Path(self.score_save_dir, f'{Path(enrol_dir).name}-{Path(test_dir).name}')
save_dir.mkdir(exist_ok=True)
sim_scores, enrol_indices, test_indices = self.compute_distances(enrol_vectors=enrol_vectors,
enrol_ids=enrol_ids,
test_vectors=test_vectors, test_ids=test_ids)
trial_scores = self.compute_trial_scores(trials=trials.keys(), enrol_indices=enrol_indices,
test_indices=test_indices, sim_scores=sim_scores,
out_file=save_dir / 'scores')
positive_scores, negative_scores = self._split_scores(trial_scores.values, trials)
del sim_scores
# Final EER computation
eer, th = EER(torch.tensor(positive_scores), torch.tensor(negative_scores))
# min_dcf, th = minDCF(torch.tensor(positive_scores), torch.tensor(negative_scores))
with open(save_dir / 'EER', 'w') as f:
f.write(str(eer))
return eer
def compute_distances(self, enrol_vectors, enrol_ids, test_vectors, test_ids):
if self.distance == 'plda':
"""Computes the Equal Error Rate give the PLDA scores"""
# Create ids, labels, and scoring list for EER evaluation
if self.plda_model_dir.exists():
self.plda = PLDAModel(train_embeddings=None, results_path=self.plda_model_dir)
else:
logger.info('Train PLDA model...')
plda_data_dir = self.plda_train_data_dir
if self.plda_anon:
plda_data_dir = Path(f'{plda_data_dir}_selected')
self.select_data_for_plda(all_data_dir=self.plda_train_data_dir,
selected_data_dir=self.model_dir.parent,
out_dir=plda_data_dir)
logger.info(f'Using data under {plda_data_dir}')
train_dict = self.extractor.extract_speakers(dataset_path=plda_data_dir, emb_level='utt')
self.plda = PLDAModel(train_embeddings=train_dict, results_path=self.plda_model_dir)
plda_score_object = self.plda.compute_distance(enrollment_vectors=enrol_vectors, enrollment_ids=enrol_ids,
trial_vectors=test_vectors, trial_ids=test_ids,
return_object=True)
sim_scores = plda_score_object.scoremat
enrol_indices = dict(zip(plda_score_object.modelset, range(len(enrol_ids))))
test_indices = dict(zip(plda_score_object.segset, range(len(test_ids))))
else:
"""Computes the Equal Error Rate give the cosine score"""
sim_scores = 1 - cosine_distances(X=enrol_vectors.cpu(), Y=test_vectors.cpu())
enrol_indices = dict(zip(enrol_ids, range(len(enrol_ids))))
test_indices = dict(zip(test_ids, range(len(test_ids))))
return sim_scores, enrol_indices, test_indices