-
Notifications
You must be signed in to change notification settings - Fork 12
/
main.py
287 lines (236 loc) · 12.9 KB
/
main.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
import argparse
import sys
import os
import numpy as np
import torch
from torch import nn
from torch import Tensor
from torch.utils.data import DataLoader
import yaml
from data_utils_rawboost import genSpoof_list,Dataset_ASVspoof2019_train,Dataset_ASVspoof2021_eval
from model import RawNet
from tensorboardX import SummaryWriter
from core_scripts.startup_config import set_random_seed
__author__ = "Hemlata Tak"
__email__ = "[email protected]"
def evaluate_accuracy(dev_loader, model, device):
num_correct = 0.0
num_total = 0.0
model.eval()
for batch_x, batch_y in dev_loader:
batch_size = batch_x.size(0)
num_total += batch_size
batch_x = batch_x.to(device)
batch_y = batch_y.view(-1).type(torch.int64).to(device)
batch_out = model(batch_x)
_, batch_pred = batch_out.max(dim=1)
num_correct += (batch_pred == batch_y).sum(dim=0).item()
return 100 * (num_correct / num_total)
def produce_evaluation_file(dataset, model, device, save_path):
data_loader = DataLoader(dataset, batch_size=128, shuffle=False, drop_last=False)
model.eval()
for batch_x,utt_id in data_loader:
fname_list = []
score_list = []
batch_size = batch_x.size(0)
batch_x = batch_x.to(device)
batch_out = model(batch_x,is_test=True)
batch_score = (batch_out[:, 1]
).data.cpu().numpy().ravel()
# add outputs
fname_list.extend(utt_id)
score_list.extend(batch_score.tolist())
with open(save_path, 'a+') as fh:
for f, cm in zip(fname_list,score_list):
fh.write('{} {}\n'.format(f, cm))
fh.close()
print('Scores saved to {}'.format(save_path))
def train_epoch(train_loader, model, lr,optim, device):
running_loss = 0
num_correct = 0.0
num_total = 0.0
ii = 0
model.train()
#set objective (Loss) functions
weight = torch.FloatTensor([0.1, 0.9]).to(device)
criterion = nn.CrossEntropyLoss(weight=weight)
for batch_x, batch_y in train_loader:
batch_size = batch_x.size(0)
num_total += batch_size
ii += 1
batch_x = batch_x.to(device)
batch_y = batch_y.view(-1).type(torch.int64).to(device)
batch_out = model(batch_x)
batch_loss = criterion(batch_out, batch_y)
_, batch_pred = batch_out.max(dim=1)
num_correct += (batch_pred == batch_y).sum(dim=0).item()
running_loss += (batch_loss.item() * batch_size)
if ii % 10 == 0:
sys.stdout.write('\r \t {:.2f}'.format(
(num_correct/num_total)*100))
optim.zero_grad()
batch_loss.backward()
optim.step()
running_loss /= num_total
train_accuracy = (num_correct/num_total)*100
return running_loss, train_accuracy
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='ASVspoof2021 baseline system')
# Dataset
parser.add_argument('--database_path', type=str, default='/your/path/to/data/ASVspoof_database/LA/', help='Change this to user\'s full directory address of LA database (ASVspoof2019- for training & development (used as validation), ASVspoof2021 for evaluation scores). We assume that all three ASVspoof 2019 LA train, LA dev and ASVspoof2021 LA eval data folders are in the same database_path directory.')
'''
% database_path/
% |- ASVspoof2021_LA_eval/flac
% |- ASVspoof2019_LA_train/flac
% |- ASVspoof2019_LA_dev/flac
'''
parser.add_argument('--protocols_path', type=str, default='/your/path/to/protocols/ASVspoof_database/', help='Change with path to user\'s LA database protocols directory address')
'''
% protocols_path/
% |- ASVspoof_LA_cm_protocols
% |- ASVspoof2021.LA.cm.eval.trl.txt
% |- ASVspoof2019.LA.cm.dev.trl.txt
% |- ASVspoof2019.LA.cm.train.trn.txt
'''
# Hyperparameters
parser.add_argument('--batch_size', type=int, default=128)
parser.add_argument('--num_epochs', type=int, default=100)
parser.add_argument('--lr', type=float, default=0.0001)
parser.add_argument('--weight_decay', type=float, default=0.0001)
parser.add_argument('--loss', type=str, default='weighted_CCE')
# model
parser.add_argument('--seed', type=int, default=1234,
help='random seed (default: 1234)')
parser.add_argument('--model_path', type=str,
default=None, help='Model checkpoint')
parser.add_argument('--comment', type=str, default=None,
help='Comment to describe the saved model')
# Auxiliary arguments
parser.add_argument('--track', type=str, default='LA',choices=['LA', 'PA','DF'], help='LA/PA/DF')
parser.add_argument('--eval_output', type=str, default=None,
help='Path to save the evaluation result')
parser.add_argument('--eval', action='store_true', default=False,
help='eval mode')
parser.add_argument('--is_eval', action='store_true', default=False,help='eval database')
parser.add_argument('--eval_part', type=int, default=0)
# backend options
parser.add_argument('--cudnn-deterministic-toggle', action='store_false', \
default=True,
help='use cudnn-deterministic? (default true)')
parser.add_argument('--cudnn-benchmark-toggle', action='store_true', \
default=False,
help='use cudnn-benchmark? (default false)')
##===================================================Rawboost data augmentation parameters======================================================================#
parser.add_argument('--algo', type=int, default=0,
help='Rawboost algos discriptions. 0: No augmentation 1: LnL_convolutive_noise, 2: ISD_additive_noise, 3: SSI_additive_noise, 4: series algo (1+2+3), \
5: series algo (1+2), 6: series algo (1+3), 7: series algo(2+3), 8: parallel algo(1,2) .[default=0]')
# LnL_convolutive_noise parameters
parser.add_argument('--nBands', type=int, default=5,
help='number of notch filters.The higher the number of bands, the more aggresive the distortions is.[default=5]')
parser.add_argument('--minF', type=int, default=20,
help='minimum centre frequency [Hz] of notch filter.[default=20] ')
parser.add_argument('--maxF', type=int, default=8000,
help='maximum centre frequency [Hz] (<sr/2) of notch filter.[default=8000]')
parser.add_argument('--minBW', type=int, default=100,
help='minimum width [Hz] of filter.[default=100] ')
parser.add_argument('--maxBW', type=int, default=1000,
help='maximum width [Hz] of filter.[default=1000] ')
parser.add_argument('--minCoeff', type=int, default=10,
help='minimum filter coefficients. More the filter coefficients more ideal the filter slope.[default=10]')
parser.add_argument('--maxCoeff', type=int, default=100,
help='maximum filter coefficients. More the filter coefficients more ideal the filter slope.[default=100]')
parser.add_argument('--minG', type=int, default=0,
help='minimum gain factor of linear component.[default=0]')
parser.add_argument('--maxG', type=int, default=0,
help='maximum gain factor of linear component.[default=0]')
parser.add_argument('--minBiasLinNonLin', type=int, default=5,
help=' minimum gain difference between linear and non-linear components.[default=5]')
parser.add_argument('--maxBiasLinNonLin', type=int, default=20,
help=' maximum gain difference between linear and non-linear components.[default=20]')
parser.add_argument('--N_f', type=int, default=5,
help='order of the (non-)linearity where N_f=1 refers only to linear components.[default=5]')
# ISD_additive_noise parameters
parser.add_argument('--P', type=int, default=10,
help='Maximum number of uniformly distributed samples in [%].[defaul=10]')
parser.add_argument('--g_sd', type=int, default=2,
help='gain parameters > 0. [default=2]')
# SSI_additive_noise parameters
parser.add_argument('--SNRmin', type=int, default=10,
help='Minimum SNR value for coloured additive noise.[defaul=10]')
parser.add_argument('--SNRmax', type=int, default=40,
help='Maximum SNR value for coloured additive noise.[defaul=40]')
##===================================================Rawboost data augmentation ======================================================================#
dir_yaml = os.path.splitext('model_config_RawNet')[0] + '.yaml'
with open(dir_yaml, 'r') as f_yaml:
parser1 = yaml.load(f_yaml)
if not os.path.exists('models'):
os.mkdir('models')
args = parser.parse_args()
#make experiment reproducible
set_random_seed(args.seed, args)
track = args.track
assert track in ['LA', 'PA','DF'], 'Invalid track given'
#database
prefix = 'ASVspoof_{}'.format(track)
prefix_2019 = 'ASVspoof2019.{}'.format(track)
prefix_2021 = 'ASVspoof2021.{}'.format(track)
#define model saving path
model_tag = 'model_{}_{}_{}_{}_{}'.format(
track, args.loss, args.num_epochs, args.batch_size, args.lr)
if args.comment:
model_tag = model_tag + '_{}'.format(args.comment)
model_save_path = os.path.join('models', model_tag)
#set model save directory
if not os.path.exists(model_save_path):
os.mkdir(model_save_path)
#GPU device
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print('Device: {}'.format(device))
#model
model = RawNet(parser1['model'], device)
nb_params = sum([param.view(-1).size()[0] for param in model.parameters()])
model =(model).to(device)
#set Adam optimizer
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr,weight_decay=args.weight_decay)
if args.model_path:
model.load_state_dict(torch.load(args.model_path,map_location=device))
print('Model loaded : {}'.format(args.model_path))
#evaluation
if args.eval:
file_eval = genSpoof_list( dir_meta = os.path.join(args.protocols_path+'{}_cm_protocols/{}.cm.eval.trl.txt'.format(prefix,prefix_2021)),is_train=False,is_eval=True)
print('no. of eval trials',len(file_eval))
eval_set=Dataset_ASVspoof2021_eval(list_IDs = file_eval,base_dir = os.path.join(args.database_path+'ASVspoof2021_{}_eval/'.format(args.track)))
produce_evaluation_file(eval_set, model, device, args.eval_output)
sys.exit(0)
# define train dataloader
d_label_trn,file_train = genSpoof_list( dir_meta = os.path.join(args.protocols_path+'{}_cm_protocols/{}.cm.train.trn.txt'.format(prefix,prefix_2019)),is_train=True,is_eval=False)
print('no. of training trials',len(file_train))
train_set=Dataset_ASVspoof2019_train(args,list_IDs = file_train,
labels = d_label_trn,
base_dir =os.path.join(args.database_path+'ASVspoof2019_{}_train/'.format(args.track)),algo=args.algo)
train_loader = DataLoader(train_set, batch_size=args.batch_size,num_workers=8, shuffle=True,drop_last = True)
del train_set,d_label_trn
# define validation dataloader
d_label_dev,file_dev = genSpoof_list( dir_meta = os.path.join(args.protocols_path+'{}_cm_protocols/{}.cm.dev.trl.txt'.format(prefix,prefix_2019)),is_train=False,is_eval=False)
print('no. of validation trials',len(file_dev))
dev_set = Dataset_ASVspoof2019_train(args,list_IDs = file_dev,
labels = d_label_dev,
base_dir = os.path.join(args.database_path+'ASVspoof2019_{}_dev/'.format(args.track)),algo=args.algo)
dev_loader = DataLoader(dev_set, batch_size=args.batch_size,num_workers=8, shuffle=False)
del dev_set,d_label_dev
# Training and validation
num_epochs = args.num_epochs
writer = SummaryWriter('logs/{}'.format(model_tag))
best_acc = 99
for epoch in range(num_epochs):
running_loss, train_accuracy = train_epoch(train_loader,model, args.lr,optimizer, device)
valid_accuracy = evaluate_accuracy(dev_loader, model, device)
writer.add_scalar('train_accuracy', train_accuracy, epoch)
writer.add_scalar('valid_accuracy', valid_accuracy, epoch)
writer.add_scalar('loss', running_loss, epoch)
print('\n{} - {} - {:.2f} - {:.2f}'.format(epoch,
running_loss, train_accuracy, valid_accuracy))
if valid_accuracy > best_acc:
print('best model find at epoch', epoch)
best_acc = max(valid_accuracy, best_acc)
torch.save(model.state_dict(), os.path.join(model_save_path, 'epoch_{}.pth'.format(epoch)))