/
train_cross_device.py
210 lines (183 loc) · 11 KB
/
train_cross_device.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
from __future__ import print_function, division
import argparse
import os
import torch
import numpy as np
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from datetime import datetime
from m3fas_utils.dataloader_fusion import face_datareader, my_transforms
from m3fas_utils.metrics import get_metrics
from arch.model import Classifier
from tqdm import tqdm
import random
import torch.nn.functional as F
import pandas as pd
os.environ['CUDA_VISIBLE_DEVICES'] = "1"
def parse_args():
parser = argparse.ArgumentParser(description='chenqi_echofas')
parser.add_argument('--lr', default=1e-4, type=float)
parser.add_argument('--save_model_epoch', default=1, type=int)
parser.add_argument('--disp_step', default=300, type=int)
parser.add_argument('--warm_start_epoch', default=0, type=int)
parser.add_argument('--epochs', default=100, type=int)
parser.add_argument('--random_seed', default=0, type=int)
parser.add_argument('--batch_size_train', default=256, type=int)
parser.add_argument('--batch_size_val', default=256, type=int)
parser.add_argument('--input_size', default=128, type=int)
parser.add_argument('--num_workers', default=8, type=int)
parser.add_argument('--loss_weight', default=0.5, type=float)
parser.add_argument('--weight_decay', default=0.00001, type=float)
parser.add_argument('--thre', default=0.5, type=float)
parser.add_argument('--save_root', default='./Training_results/Cross_device_fusion/', type=str)
parser.add_argument('--root_path', default='./', type=str)
parser.add_argument('--model_name', default="2note", type=str) # Please change the phone device accordingly (note, s9, s21, xiaomi).
parser.add_argument('--train_csv', default="./csv/Cross_device_csv/2note_train.csv", type=str) # Please change the phone device accordingly (note, s9, s21, xiaomi).
parser.add_argument('--val_csv', default="./csv/Cross_device_csv/2note_val.csv", type=str) # Please change the phone device accordingly (note, s9, s21, xiaomi).
parser.add_argument('--test_csv', default="./csv/Cross_device_csv/2note_test.csv", type=str) # Please change the phone device accordingly (note, s9, s21, xiaomi).
return parser.parse_args()
def fix_seed(seed):
torch.cuda.manual_seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
def Validation(model, dataloader, args, thre, facenum):
model.eval()
GT = np.zeros((facenum,), np.int)
PRED_f = np.zeros((facenum,), np.float)
PRED_v = np.zeros((facenum,), np.float)
PRED_a = np.zeros((facenum,), np.float)
for num, data in enumerate(tqdm(dataloader)):
with torch.no_grad():
faces = data['faces'].to('cuda')
spects = data['spects'].to('cuda')
labels = data['labels'].to('cuda')
logits_f, logits_v, logits_a = model(faces, spects)
pred_score_f = torch.nn.functional.softmax(logits_f, 1)
pred_score_v = torch.nn.functional.softmax(logits_v, 1)
pred_score_a = torch.nn.functional.softmax(logits_a, 1)
GT[num * args.batch_size_val: (num*args.batch_size_val+faces.size(0))]= labels.cpu().numpy()
PRED_f[num * args.batch_size_val: (num*args.batch_size_val+faces.size(0))] = pred_score_f[:,1].cpu().numpy()
PRED_v[num * args.batch_size_val: (num*args.batch_size_val+faces.size(0))] = pred_score_v[:,1].cpu().numpy()
PRED_a[num * args.batch_size_val: (num*args.batch_size_val+faces.size(0))] = pred_score_a[:,1].cpu().numpy()
acc_f, auc_f, hter_f, eer_f = get_metrics(PRED_f, GT, thre)
acc_v, auc_v, hter_v, eer_v = get_metrics(PRED_v, GT, thre)
acc_a, auc_a, hter_a, eer_a = get_metrics(PRED_a, GT, thre)
return auc_f, hter_f, eer_f, acc_f, auc_v, hter_v, eer_v, acc_v, auc_a, hter_a, eer_a, acc_a
def train(args, model):
avg_train_loss_list = np.array([])
train_dataset = face_datareader(csv_file=args.train_csv, transform=my_transforms(size=args.input_size))
training_dataloader = DataLoader(train_dataset, batch_size=args.batch_size_train, shuffle=True, num_workers=args.num_workers, drop_last=False, pin_memory=True)
val_dataset = face_datareader(csv_file=args.val_csv, transform=my_transforms(size=args.input_size))
val_dataloader = DataLoader(val_dataset, batch_size=args.batch_size_val, shuffle=True, num_workers=args.num_workers, drop_last=False, pin_memory=True)
test_dataset = face_datareader(csv_file=args.test_csv, transform=my_transforms(size=args.input_size))
test_dataloader = DataLoader(test_dataset, batch_size=args.batch_size_val, shuffle=True, num_workers=args.num_workers, drop_last=False, pin_memory=True)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
Num_val_faces = len(pd.read_csv(args.val_csv, header=None))
Num_test_faces = len(pd.read_csv(args.test_csv, header=None))
# result folder
res_folder_name = args.save_root + args.model_name
if not os.path.exists(res_folder_name):
os.makedirs(res_folder_name)
os.mkdir(res_folder_name + '/ckpt/')
else:
print("WARNING: RESULT PATH ALREADY EXISTED -> " + res_folder_name)
print('find models here: ', res_folder_name)
writer = SummaryWriter(res_folder_name)
f1 = open(res_folder_name + "/training_log.csv", 'a+')
# training
steps_per_epoch = len(training_dataloader)
Best_HTER = 1.0
for epoch in range(args.warm_start_epoch, args.epochs):
step_loss = np.zeros(steps_per_epoch, dtype=np.float)
step_loss_f = np.zeros(steps_per_epoch, dtype=np.float)
step_loss_v = np.zeros(steps_per_epoch, dtype=np.float)
step_loss_a = np.zeros(steps_per_epoch, dtype=np.float)
# scheduler.step()
for step, data in enumerate(tqdm(training_dataloader)):
model.train()
optimizer.zero_grad()
faces = data['faces'].to('cuda')
spects = data['spects'].to('cuda')
labels = data['labels'].to('cuda')
logits_f, logits_v, logits_a = model(faces, spects)
loss_f = F.cross_entropy(logits_f, labels)
loss_v = F.cross_entropy(logits_v, labels)
loss_a = F.cross_entropy(logits_a, labels)
loss = loss_f + args.loss_weight*(loss_v + loss_a)
step_loss[step] = loss
step_loss_f[step] = loss_f
step_loss_v[step] = loss_v
step_loss_a[step] = loss_a
loss.backward()
optimizer.step()
Global_step = epoch * steps_per_epoch + (step + 1)
if Global_step % args.disp_step == 0:
avg_loss = np.mean(step_loss[(step + 1) - args.disp_step: (step + 1)])
avg_loss_f = np.mean(step_loss_f[(step + 1) - args.disp_step: (step + 1)])
avg_loss_v = np.mean(step_loss_v[(step + 1) - args.disp_step: (step + 1)])
avg_loss_a = np.mean(step_loss_a[(step + 1) - args.disp_step: (step + 1)])
now_time = datetime.now().strftime("%d/%m/%Y %H:%M:%S")
step_log_msg = '[%s] Epoch: %d/%d | Global_step: %d | Avg_loss: %f | Avg_loss_f: %f | Avg_loss_v: %f | Avg_loss_a: %f ' % (now_time, epoch + 1, args.epochs, Global_step, avg_loss, avg_loss_f, avg_loss_v, avg_loss_a)
writer.add_scalar('Loss/train', avg_loss, Global_step)
print('\n', step_log_msg)
if (epoch+1) % args.save_model_epoch == 0:
now_time = datetime.now().strftime("%d/%m/%Y %H:%M:%S")
avg_train_loss = np.mean(step_loss[(step + 1) - args.disp_step: (step + 1)])
avg_train_loss_list = np.append(avg_train_loss_list, avg_train_loss)
log_msg = '[%s] Epoch: %d/%d | 1/10 average epoch loss: %f' % (now_time, epoch + 1, args.epochs, avg_train_loss)
print('\n', log_msg)
f1.write(log_msg)
f1.write('\n')
# validation
print('Validating...')
AUC_f, HTER_f, EER_f, ACC_f, AUC_v, HTER_v, EER_v, ACC_v, AUC_a, HTER_a, EER_a, ACC_a = Validation(model, val_dataloader, args, args.thre, Num_val_faces)
val_msg_f = '[%s] Epoch: %d/%d | Global_step: %d | AUC_f: %f | HTER_f: %f | EER_f: %f | ACC_f: %f' % (now_time, epoch + 1, args.epochs, Global_step, AUC_f, HTER_f, EER_f, ACC_f)
print('\n', val_msg_f)
val_msg_v = '[%s] Epoch: %d/%d | Global_step: %d | AUC_v: %f | HTER_v: %f | EER_v: %f | ACC_v: %f' % (now_time, epoch + 1, args.epochs, Global_step, AUC_v, HTER_v, EER_v, ACC_v)
print('\n', val_msg_v)
val_msg_a = '[%s] Epoch: %d/%d | Global_step: %d | AUC_a: %f | HTER_a: %f | EER_a: %f | ACC_a: %f' % (now_time, epoch + 1, args.epochs, Global_step, AUC_a, HTER_a, EER_a, ACC_a)
print('\n', val_msg_a)
f1.write(val_msg_f)
f1.write('\n')
f1.write(val_msg_v)
f1.write('\n')
f1.write(val_msg_a)
f1.write('\n')
# Here we pick the last checkpoint as the final ckpt. You can also try some alternative strategies.
#save model
if not HTER_f > Best_HTER:
# Best_HTER = HTER_f
torch.save(model.state_dict(), res_folder_name + '/ckpt/' + 'best.pth')
np.save(res_folder_name + '/avg_train_loss_list.np', avg_train_loss_list)
cur_learning_rate = [param_group['lr'] for param_group in optimizer.param_groups]
print('Saved model. lr %f' % cur_learning_rate[0])
f1.write('Saved model. lr %f' % cur_learning_rate[0])
f1.write('\n')
print('Testing...')
AUC_f, HTER_f, EER_f, ACC_f, AUC_v, HTER_v, EER_v, ACC_v, AUC_a, HTER_a, EER_a, ACC_a = Validation(model, test_dataloader, args, args.thre, Num_test_faces)
test_msg_f = '[%s] Epoch: %d/%d | Global_step: %d | AUC_f: %f | HTER_f: %f | EER_f: %f | ACC_f: %f' % (now_time, epoch + 1, args.epochs, Global_step, AUC_f, HTER_f, EER_f, ACC_f)
print('\n', test_msg_f)
test_msg_v = '[%s] Epoch: %d/%d | Global_step: %d | AUC_v: %f | HTER_v: %f | EER_v: %f | ACC_v: %f' % (now_time, epoch + 1, args.epochs, Global_step, AUC_v, HTER_v, EER_v, ACC_v)
print('\n', test_msg_v)
test_msg_a = '[%s] Epoch: %d/%d | Global_step: %d | AUC_a: %f | HTER_a: %f | EER_a: %f | ACC_a: %f' % (now_time, epoch + 1, args.epochs, Global_step, AUC_a, HTER_a, EER_a, ACC_a)
print('\n', test_msg_a)
f1.write(test_msg_f)
f1.write('\n')
f1.write(test_msg_v)
f1.write('\n')
f1.write(test_msg_a)
f1.write('\n')
f1.close()
if __name__ == '__main__':
args = parse_args()
print(args)
if args.random_seed is not None:
fix_seed(args.random_seed)
model = Classifier()
model = model.to('cuda')
train(args, model)