-
Notifications
You must be signed in to change notification settings - Fork 19
/
save_embeddings.py
executable file
·193 lines (149 loc) · 5.57 KB
/
save_embeddings.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
import argparse
import json
import os
import pickle as pk
from pathlib import Path
import numpy as np
import torch
from torch.utils.data import DataLoader
from tqdm import tqdm
from aasist.data_utils import Dataset_ASVspoof2019_devNeval
from aasist.models.AASIST import Model as AASISTModel
from ECAPATDNN.model import ECAPA_TDNN
from utils import load_parameters
# list of dataset partitions
SET_PARTITION = ["trn", "dev", "eval"]
# list of countermeasure(CM) protocols
SET_CM_PROTOCOL = {
"trn": "protocols/ASVspoof2019.LA.cm.train.trn.txt",
"dev": "protocols/ASVspoof2019.LA.cm.dev.trl.txt",
"eval": "protocols/ASVspoof2019.LA.cm.eval.trl.txt",
}
# directories of each dataset partition
SET_DIR = {
"trn": "./LA/ASVspoof2019_LA_train/",
"dev": "./LA/ASVspoof2019_LA_dev/",
"eval": "./LA/ASVspoof2019_LA_eval/",
}
# enrolment data list for speaker model calculation
# each speaker model comprises multiple enrolment utterances
SET_TRN = {
"dev": [
"./LA/ASVspoof2019_LA_asv_protocols/ASVspoof2019.LA.asv.dev.female.trn.txt",
"./LA/ASVspoof2019_LA_asv_protocols/ASVspoof2019.LA.asv.dev.male.trn.txt",
],
"eval": [
"./LA/ASVspoof2019_LA_asv_protocols/ASVspoof2019.LA.asv.eval.female.trn.txt",
"./LA/ASVspoof2019_LA_asv_protocols/ASVspoof2019.LA.asv.eval.male.trn.txt",
],
}
def save_embeddings(
set_name, cm_embd_ext, asv_embd_ext, device
):
meta_lines = open(SET_CM_PROTOCOL[set_name], "r").readlines()
utt2spk = {}
utt_list = []
for line in meta_lines:
tmp = line.strip().split(" ")
spk = tmp[0]
utt = tmp[1]
if utt in utt2spk:
print("Duplicated utt error", utt)
utt2spk[utt] = spk
utt_list.append(utt)
base_dir = SET_DIR[set_name]
dataset = Dataset_ASVspoof2019_devNeval(utt_list, Path(base_dir))
loader = DataLoader(
dataset, batch_size=30, shuffle=False, drop_last=False, pin_memory=True
)
cm_emb_dic = {}
asv_emb_dic = {}
print("Getting embeddings from set %s..." % (set_name))
for batch_x, key in tqdm(loader):
batch_x = batch_x.to(device)
with torch.no_grad():
batch_cm_emb, _ = cm_embd_ext(batch_x)
batch_cm_emb = batch_cm_emb.detach().cpu().numpy()
batch_asv_emb = asv_embd_ext(batch_x, aug=False).detach().cpu().numpy()
for k, cm_emb, asv_emb in zip(key, batch_cm_emb, batch_asv_emb):
cm_emb_dic[k] = cm_emb
asv_emb_dic[k] = asv_emb
os.makedirs("embeddings", exist_ok=True)
with open( "embeddings/cm_embd_%s.pk" % (set_name), "wb") as f:
pk.dump(cm_emb_dic, f)
with open("embeddings/asv_embd_%s.pk" % (set_name), "wb") as f:
pk.dump(asv_emb_dic, f)
def save_models(set_name, asv_embd_ext, device):
utt2spk = {}
utt_list = []
for trn in SET_TRN[set_name]:
meta_lines = open(trn, "r").readlines()
for line in meta_lines:
tmp = line.strip().split(" ")
spk = tmp[0]
utts = tmp[1].split(",")
for utt in utts:
if utt in utt2spk:
print("Duplicated utt error", utt)
utt2spk[utt] = spk
utt_list.append(utt)
base_dir = SET_DIR[set_name]
dataset = Dataset_ASVspoof2019_devNeval(utt_list, Path(base_dir))
loader = DataLoader(
dataset, batch_size=30, shuffle=False, drop_last=False, pin_memory=True
)
asv_emb_dic = {}
print("Getting embedgins from set %s..." % (set_name))
for batch_x, key in tqdm(loader):
batch_x = batch_x.to(device)
with torch.no_grad():
batch_asv_emb = asv_embd_ext(batch_x, aug=False).detach().cpu().numpy()
for k, asv_emb in zip(key, batch_asv_emb):
utt = k
spk = utt2spk[utt]
if spk not in asv_emb_dic:
asv_emb_dic[spk] = []
asv_emb_dic[spk].append(asv_emb)
for spk in asv_emb_dic:
asv_emb_dic[spk] = np.mean(asv_emb_dic[spk], axis=0)
with open("embeddings/spk_model.pk_%s" % (set_name), "wb") as f:
pk.dump(asv_emb_dic, f)
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"-aasist_config", type=str, default="./aasist/config/AASIST.conf"
)
parser.add_argument(
"-aasist_weight", type=str, default="./aasist/models/weights/AASIST.pth"
)
parser.add_argument(
"-ecapa_weight", type=str, default="./ECAPATDNN/exps/pretrain.model"
)
return parser.parse_args()
def main():
args = get_args()
device = "cuda" if torch.cuda.is_available() else "cpu"
print("Device: {}".format(device))
with open(args.aasist_config, "r") as f_json:
config = json.loads(f_json.read())
model_config = config["model_config"]
cm_embd_ext = AASISTModel(model_config).to(device)
load_parameters(cm_embd_ext.state_dict(), args.aasist_weight)
cm_embd_ext.to(device)
cm_embd_ext.eval()
asv_embd_ext = ECAPA_TDNN(C=1024)
load_parameters(asv_embd_ext.state_dict(), args.ecapa_weight)
asv_embd_ext.to(device)
asv_embd_ext.eval()
for set_name in SET_PARTITION:
save_embeddings(
set_name,
cm_embd_ext,
asv_embd_ext,
device,
)
if set_name == "trn":
continue
save_models(set_name, asv_embd_ext, device)
if __name__ == "__main__":
main()