-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
136 lines (79 loc) · 3.4 KB
/
train.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
import torch
from torch import optim
from torchvision import datasets, transforms
from tensorboardX import SummaryWriter
import pandas as pd
from torch.nn import CrossEntropyLoss
import os
import matplotlib.pyplot as plt
from model import *
from tqdm import tqdm
import cv2
import numpy as np
from datetime import datetime
from evalsource import evaluate
model_path = "../model/"
log_dir = "../LOG/"
name = str(datetime.now())[:-10].replace(' ','-').replace(':','-')
writer = SummaryWriter(log_dir)
os.makedirs(os.path.dirname(model_path), exist_ok=True)
os.makedirs(os.path.dirname(log_dir+name+'/'), exist_ok=True)
###### PREPARE DATA ######
transform = transforms.Compose([transforms.Resize(112), transforms.CenterCrop(112),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(), transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])
dataset = datasets.ImageFolder('../imgs/', transform=transform)
class_num = dataset[-1][1] + 1
dataloader = torch.utils.data.DataLoader(dataset, batch_size=32, shuffle=True, num_workers=2)
model = MobileFaceNet(512).to(torch.device("cuda:0"))
arc = Arcface(embedding_size=512, classnum=class_num).to(torch.device("cuda:0"))
paras_only_bn, paras_wo_bn = separate_bn_paras(model)
optimizer = optim.SGD([
{'params': paras_wo_bn[:-1], 'weight_decay': 4e-5},
{'params': [paras_wo_bn[-1]] + [arc.kernel], 'weight_decay': 4e-4},
{'params': paras_only_bn}
], lr = 1e-3, momentum = 0.9)
###### LOAD MODEL ######
model.load_state_dict(torch.load('../model/2020-11-15-01-10/model_accuracy:0.6235.pth'))
arc.load_state_dict(torch.load('../model/2020-11-15-01-10/arc_accuracy:0.6235.pth'))
optimizer.optimizer.load_state_dict(torch.load('../model/2020-11-15-01-10/optimizer_accuracy:0.6235.pth'))
###### LOAD LFW IMG ######
lfw, lfw_issame = get_val_pair('data', 'lfw')
acc, std, best = evaluate(model=model, carray=lfw, issame=lfw_issame)
print("accuracy: %1.3f" %(acc*100))
model.train()
f = open('step.txt', 'r')
global_step = int(f.readline())
epoch = int(f.readline())
while epoch < 500:
for img, label in iter(dataloader):
img = img.to(torch.device("cuda:0"))
label = label.to(torch.device("cuda:0"))
optimizer.zero_grad()
embedding = model(img)
theta = arc(embedding, label)
loss = CrossEntropyLoss()(theta, label)
loss.backward()
optimizer.step()
if global_step%100 == 0:
f = open('step.txt', 'w')
f.write(str(global_step) + '\n' + str(epoch))
f.close()
writer.add_scalar('loss', loss, global_step)
print("Global step: %.d ==== Epoch: %d ==== Loss: %.3f" % (global_step, epoch, loss))
if global_step%10000 == 0 and global_step != 0:
acc, std, best = evaluate(model=model, carray=lfw, issame=lfw_issame)
acc = np.round(acc*100, 3)
std = np.round(std*100, 3)
print("Accuracy: {}+-{}".format(acc, std))
print("Best threshold: {}".format(best))
name = str(datetime.now())[:-10].replace(' ','-').replace(':','-')
model_save = os.path.join(model_path, name +'/')
os.makedirs(os.path.dirname(model_save), exist_ok=True)
torch.save(model.state_dict(), model_save + 'model_accuracy:{}.pth'.format(acc))
torch.save(arc.state_dict(), model_save + 'arc_accuracy:{}.pth'.format(acc))
torch.save(optimizer.state_dict(), model_save + 'optimizer_accuracy:{}.pth'.format(acc))
writer.add_scalar('accuracy', acc, global_step)
model.train()
global_step += 1
epoch += 1