-
Notifications
You must be signed in to change notification settings - Fork 0
/
csl_train_slf.py
146 lines (112 loc) · 5.7 KB
/
csl_train_slf.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
import os
import warnings
import random
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from sklearn.metrics import roc_auc_score
from torch.utils.data import DataLoader
from tqdm import *
from dataset.dataset import MyDataset
from model.model import Contrastive
from model.loss import ContrastiveLoss
warnings.filterwarnings("ignore")
import yaml
def seed_anything(seed_value):
np.random.seed(seed_value)
random.seed(seed_value)
os.environ['PYTHONHASHSEED'] = str(seed_value) # 为了禁止hash随机化，使得实验可复现。
torch.manual_seed(seed_value) # 为CPU设置随机种子
torch.cuda.manual_seed(seed_value) # 为当前GPU设置随机种子（只用一块GPU）
torch.cuda.manual_seed_all(seed_value) # 为所有GPU设置随机种子（多块GPU）
torch.backends.cudnn.deterministic = True
if __name__ == '__main__':
config_path = './config/csl/train_slf.yaml'
with open(config_path,'r',encoding='utf-8') as f:
configs = yaml.load(f,Loader=yaml.FullLoader)
print(f'EXP Settings: ')
for k,v in configs.items():
print(f'{k}: {v}')
print(f'*'*30)
seed_anything(configs['seed'])
# 创建模型保存路径
if not os.path.exists(os.path.join('./weight','csl',configs['name'])):
os.mkdir(os.path.join('./weight','csl',configs['name']))
# 创建结果保存路径
if not os.path.exists(os.path.join('./output','csl','train',configs['name'])):
os.mkdir(os.path.join('./output','csl','train',configs['name']))
# 将配置文件保存到模型保存路径
with open(os.path.join('./weight','csl',configs['name'],'train.yaml'),'w') as f:
f.write(yaml.dump(configs,allow_unicode=True))
# 初始化
train_dataset = MyDataset(root=configs['train_data_toot'],
real_sample_only=configs['real_sample_only'],
is_train=configs['is_train'],
pos_neg_rate=configs['pos_neg_rate'],
window_len=configs['window_len'],
n_extracts=configs['n_extracts'])
train_loader = DataLoader(train_dataset, batch_size = configs['batch_size'], shuffle=True, num_workers = configs['num_workers'])
val_dataset = MyDataset(root=configs['val_data_root'],
real_sample_only=False,
is_train=False,
pos_neg_rate=10,
window_len=configs['window_len'],
n_extracts=configs['n_extracts'])
val_loader = DataLoader(val_dataset, batch_size = configs['batch_size'], shuffle=True, num_workers = configs['num_workers'])
model = Contrastive(window_len=configs['window_len'],
fix_backbone=configs['fix_backbone'],
device=configs['device'],
face_backbone=configs['face_backbone'])
model.to(configs['device'])
if configs['pretrained_ckpt'] != None and configs['pretrained_ckpt'] != False:
model.load_state_dict(torch.load(configs['pretrained_ckpt']))
criterion = ContrastiveLoss(margin=configs['margin'],device=configs['device'])
opt = optim.Adam(model.parameters(),lr=configs['lr'])
train_loss_curve = []
for epoch in range(configs['epochs']):
total_loss = 0.0
model.train()
for _,(face, lip, landmark, label, face_label) in tqdm(enumerate(train_loader),desc = f'on training epoch {epoch}'):
opt.zero_grad()
d = model(lip,landmark)
loss = criterion(d,label)
loss.backward()
opt.step()
total_loss += loss.item()
train_loss_curve.append(total_loss)
print(f"Epoch {epoch + 1}/{configs['epochs']}, Loss: {total_loss / len(train_loader)}")
if epoch % configs['val_gap'] == 0 or epoch >= configs['epochs'] - 1:
model.eval()
label_list = []
distance_list = []
with torch.no_grad():
for _,(face, lip, landmark, label, face_label) in tqdm(enumerate(val_loader),desc = f'on testing epoch'):
label = label.float().to(configs['device'])
distance = model(lip,landmark)
label_list += label.cpu().float().tolist()
distance_list += distance.cpu().float().tolist()
# draw distances
real_distance = []
fake_distance = []
for i in range(len(label_list)):
if label_list[i] > 0.5:
fake_distance.append(distance_list[i])
else:
real_distance.append(distance_list[i])
real_distance = np.array(real_distance)
fake_distance = np.array(fake_distance)
x1 = np.array([i for i in range(len(real_distance))])
x2 = np.array([i for i in range(len(fake_distance))])
plt.scatter(x1,real_distance,c='r')
plt.scatter(x2,fake_distance,c='b')
plt.legend(['real_d','fake_d'])
plt.savefig(os.path.join('./output','csl','train',configs['name'],f'ep{epoch}_distance.png'))
# plt.savefig(f'{name}_distance.png')
plt.clf()
# 保存模型
torch.save(model.state_dict(),os.path.join('./weight','csl',configs['name'],f'ep{epoch}.pth'))
plt.plot(train_loss_curve)
plt.savefig(os.path.join('./output','csl','train',configs['name'],f'loss_curve.png'))
plt.clf()