本文总结如何构建insightface的数据集，我们以LFW数据集与CFP数据集为例展开实验，读者也可也用自己的数据集。
1数据对齐
在这之前你应该准备好你的数据，可以爬虫爬来，也可以自己找，这里我们采用LFW官方数据集，你的数据文件目录应该如下：
–class1 --pic1.jpg
-------------pic2.jpg
-------------pic3.jpg
–class2 --pic1.jpg
-------------pic2.jpg
给出LFW数据的地址：
lfw数据
http://vis-www.cs.umass.edu/lfw/
我下载的是这个原图像https://drive.google.com/file/d/1p1wjaqpTh_5RHfJu4vUh8JJCdKwYMHCp/view?usp=sharing
数据集有了之后，我们这里做的是人脸识别，所以人脸检测与对齐部分我们直接跑模型即可。
datasets/lfwdata下存放原始数据（lfwdata下分目录存放每一类的数据，每个人一个文件夹，里面存图。）
跳转到src/align，这里提供了人脸对齐的代码，检测加对齐MTCNN；运行align_lfw.py。两个参数input为原始数据文件夹；output是你要存放对齐后的数据。
python3 align_lfw.py --input-dir ../../datasets/lfwdata --output-dir ../../datasets/lfw2
1
附录：运行./src/align/align_lfw.py遇到的错误
1>tf
pip3 install tensorflow-gpu -i https://pypi.douban.com/simple --user
2>Question:ValueError: Object arrays cannot be loaded when allow_pickle=False
Answer:pip3 install numpy==1.16.1 -i https://pypi.douban.com/simple --user
3>Question:AttributeError: module ‘scipy.misc’ has no attribute ‘imread’
Answer:pip3 install scipy==1.1.0 -i https://pypi.douban.com/simple --user
PS：以下为CFP-FP不同的地方，读者如果构建自己的数据集或者LFW数据集可以不看。
注意：这里cfp-fp数据集的对齐和lfw不一样，原因在于目录层级不一样，cfp-fp数据目录下每个person还分正脸侧脸，所以在遍历时不太一致，需要深度两层遍历，我们又写了align_cfp.py文件。
这里我们复制一份align_lfw.py到align_cfp.py。然后改main函数里
1、数据路径遍历方式变为ytf，这个你去看face_image有定义
dataset = face_image.get_dataset('ytf', args.input_dir)
print('dataset size', 'lfw', len(dataset))
1
2
2、113行改一下生成新的路径存储图片这里
img = img[:,:,0:3]
_paths = fimage.image_path.split('/')
a,b,c = _paths[-3], _paths[-2],_paths[-1]
1
2
3
这样就可以两层遍历找所有的图片了！
2生成list文件
所以我们找了个程序附录2代码来生成list。
…/…/datasets/lfw/train.lst输出目录最后train.lst代表list的名字。（这里我们就用lfw代表最后存储lst、rec等文件的位置；）
…/…/datasets/lfw2对齐后的图片目录
python3 generatelst.py --dataset-dir ../../datasets/lfw2 --list-file ../../datasets/lfw/train.lst --img-ext '.jpg'
1
注意：这里生成的lst格式如下：
1 path/Adam_Brody/Adam_Brody_277.png 25
其中1代表对齐，25代表class label 中间是地址；需要用\t表示tab键，不能用空格间隔。
3生成rec &idx 文件（依托于list）
生成lst文件后，直接用face2rec2文件生成rec与idx，用python3存在一些问题；可以用python2
python face2rec2.py ../../datasets/lfw/
1
或者修改face2rec2代码，修改image_encode部分pack原来是’ '改为加一个b
header = mx.recordio.IRHeader(item.flag, item.label, item.id, 0)
#print('write', item.flag, item.id, item.label)
s = mx.recordio.pack(header,b'')
q_out.put((i, s, oitem))
1
2
3
4
然后在运行python3即可
python3 face2rec2.py ../../datasets/lfw/
1
4创建property配置文件
上述完成了训练所需的rec与dix文件的生成，这里还需要创建一个配置文件，直接创建一个名为property的文件，没有后缀
写1000,112,112代表ID数量,尺寸,尺寸
1000,112,112
1
目前datasets/lfw/目录下存在train.idx train.lst train.rec property
5 创建pair文件
为了做测试我们需要生成验证集用的bin文件，bin文件生成前需要做pair文件，就是一对一对的数据，每一行分别是
图A的目录 空格 图B的目录 空格 标志0/1（代表两张图类别一致否）
对于LFW和CFP这些官方的数据，直接利用官方提供的pair文件即可。我们自己构建的数据，可以利用下面我写的一个脚本简单生成。
利用generate_image_pairs.py（源文件有问题，已修改）
稍后上传，附录1有
…/…/datasets/lfw2对齐图像目录
…/…/datasets/lfw/train.txt存放txt
3000要多少个正样本数据，会同时生成同样负样本；
python3 generate_image_pairs.py --data-dir ../../datasets/lfw2 --outputtxt ../../datasets/lfw/train.txt --num-samepairs 3000
1
PS:CFP可以往后看，非CFP可以绕过了！
注意：这里生成pairs的方法不太好，数据集给了一些标准的pairs文件，我们可以写一个脚本取解读，具体如下：
lfw在insightface里面有pair.txt
cfp没有，只有一组FP对，需要我们自己写个脚本，这里我写好了放到附录3中。我们依然放到src/data下cfp_make_bin.py
cfp给了这个目录有一组对
前面对齐后数据存放在cfpdataimg里面，但是Protocol没有拷过来，我们手动复制过来。
python3 cfp_make_bin.py --data-dir …/…/datasets/cfpdataimg --output …/…/datasets/cfp/cfp.bin
6 生成验证集bin文件
完事后利用/src/data/下的 lfw2pack.py生成bin文件
但是有点问题，需要修改下，参考这篇博客https://blog.csdn.net/hanjiangxue_wei/article/details/86566348
修改lfw2pack.py中19行，打#的为更改的，改为两个参数，一个是txt读出来的列表，另一个是总数量。
注意：cfp跳过就可以了
import mxnet as mx
from mxnet import ndarray as nd
import argparse
import pickle
import sys
import os
sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'eval'))
import lfw
parser = argparse.ArgumentParser(description='Package LFW images')
# general
parser.add_argument('--data-dir', default='', help='')
parser.add_argument('--image-size', type=str, default='112,112', help='')
parser.add_argument('--output', default='', help='path to save.')
parser.add_argument('--num-samepairs',default=100)
args = parser.parse_args()
lfw_dir = args.data_dir
image_size = [int(x) for x in args.image_size.split(',')]
lfw_pairs = lfw.read_pairs(os.path.join(lfw_dir, 'train.txt'))
print(lfw_pairs)
lfw_paths, issame_list = lfw.get_paths(lfw_pairs,int(args.num_samepairs)+1)#, 'jpg')
lfw_bins = []
#lfw_data = nd.empty((len(lfw_paths), 3, image_size[0], image_size[1]))
print(len(issame_list))
i = 0
for path in lfw_paths:
with open(path, 'rb') as fin:
_bin = fin.read()
lfw_bins.append(_bin)
#img = mx.image.imdecode(_bin)
#img = nd.transpose(img, axes=(2, 0, 1))
#lfw_data[i][:] = img
i+=1
if i%1000==0:
print('loading lfw', i)
with open(args.output, 'wb') as f:
pickle.dump((lfw_bins, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
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
对应的get_paths这个文件存在src\eval\lfw.py下，把他也改了
def get_paths(pairs, same_pairs):
nrof_skipped_pairs = 0
path_list = []
issame_list = []
cnt = 1
for pair in pairs:
path0 = pair[0]
path1 = pair[1]
if cnt < same_pairs:
issame = True
else:
issame = False
if os.path.exists(path0) and os.path.exists(path1): # Only add the pair if both paths exist
path_list += (path0,path1)
issame_list.append(issame)
else:
print('not exists', path0, path1)
nrof_skipped_pairs += 1
cnt += 1
if nrof_skipped_pairs>0:
print('Skipped %d image pairs' % nrof_skipped_pairs)
return path_list, issame_list
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
之后再运行
python3 lfw2pack.py --data-dir ../../datasets/lfw --output ../../datasets/lfw/lfw.bin --num-samepairs 300
1
2
附录1 pair文件生成代码
generate_image_pairs.py、
# coding:utf-8
import sys
import os
import random
import time
import itertools
import pdb
import argparse
#src = '../../datasets/lfw2'
#dst = open('../../datasets/lfw/train.txt', 'a')
parser = argparse.ArgumentParser(description='generate image pairs')
# general
parser.add_argument('--data-dir', default='', help='')
parser.add_argument('--outputtxt', default='', help='path to save.')
parser.add_argument('--num-samepairs',default=100)
args = parser.parse_args()
cnt = 0
same_list = []
diff_list = []
list1 = []
list2 = []
folders_1 = os.listdir(args.data_dir)
dst = open(args.outputtxt, 'a')
count = 0
dst.writelines('\n')
# 产生相同的图像对
for folder in folders_1:
sublist = []
same_list = []
imgs = os.listdir(os.path.join(args.data_dir, folder))
for img in imgs:
img_root_path = os.path.join(args.data_dir, folder, img)
sublist.append(img_root_path)
list1.append(img_root_path)
for item in itertools.combinations(sublist, 2):
for name in item:
same_list.append(name)
if len(same_list) > 10 and len(same_list) < 13:
for j in range(0, len(same_list), 2):
if count < int(args.num_samepairs):#数量可以修改
dst.writelines(same_list[j] + ' ' + same_list[j+1]+ ' ' + '1' + '\n')
count += 1
if count >= int(args.num_samepairs):
break
list2 = list1.copy()
# 产生不同的图像对
diff = 0
print(count)
# 如果不同的图像对远远小于相同的图像对，则继续重复产生，直到两者相差很小
while True:
random.seed(time.time() * 100000 % 10000)
random.shuffle(list2)
for p in range(0, len(list2) - 1, 2):
if list2[p] != list2[p + 1]:
dst.writelines(list2[p] + ' ' + list2[p + 1] + ' ' + '0'+ '\n')
diff += 1
if diff >= count:
break
#print(diff)
if diff < count:
#print('--')
continue
else:
break
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
附录2 产生lst文件代码
import os
import random
import argparse
class PairGenerator:
def __init__(self, data_dir, pairs_filepath, img_ext):
"""
Parameter data_dir, is your data directory.
Parameter pairs_filepath, where is the pairs.txt that belongs to.
Parameter img_ext, is the image data extension for all of your image data.
"""
self.data_dir = data_dir
self.pairs_filepath = pairs_filepath
self.img_ext = img_ext
# splitting the database content into 10 random sets
def split_to_10(self):
folders = []
cnt = 0
for name in os.listdir(self.data_dir):
folders.append(name)
folders = sorted(folders) # sorting names in abc order
a = []
# names of folders - e.g. Talgat Bigeldinov, Kairat Nurtas, etc.
for name in folders:
# f = open(self.pairs_filepath, 'a+')
# looping through image files in one folder
for file in os.listdir(self.data_dir + '/' + name):
# a.append(data_dir + name + '/' + file)
a.append(name)
cnt = cnt + 1
cnt = cnt + 1
random.shuffle(a)
# splitting the database content into 10 random sets
def write_similar(self, lst):
f = open(self.pairs_filepath, 'a+')
for i in range(20):
left = random.choice(lst)
right = random.choice(lst)
f.write(left + '\t' + right + '\t' + '1\n')
# writing 1 IMAGE_PATH LABEL like insightface lst file needs
def write_item_label(self):
cnt = 0
for name in os.listdir(self.data_dir):
if name == ".DS_Store":
continue
# print(name)
a = []
f = open(self.pairs_filepath, 'a+')
for file in os.listdir(self.data_dir + '/' + name):
if file == ".DS_Store":
continue
a.append(data_dir + '/' + name + '/' + file)
f.write(str(1) + '\t' + data_dir + '/' + name + '/' + file + '\t' + str(cnt) + '\n')
cnt = cnt + 1
# writing 1 IMAGE_PATH LABEL like insightface lst file needs in alphabetic order
def write_item_label_abc(self):
cnt = 0
names = []
for name in os.listdir(self.data_dir):
names.append(name)
names = sorted(names)
for name in names:
print(name)
a = []
f = open(self.pairs_filepath, 'a+')
for file in os.listdir(self.data_dir + '/' + name):
if file == ".DS_Store":
continue
a.append(data_dir + '/' + name + '/' + file)
f.write(str(1) + '\t' + data_dir + '/' + name + '/' + file + '\t' + str(cnt) + '\n')
cnt = cnt + 1
def write_different(self, lst1, lst2):
f = open(self.pairs_filepath, 'a+')
for i in range(500):
left = random.choice(lst1)
right = random.choice(lst2)
f.write(left + '\t' + right + '\t' + '0\n')
f.close()
def generate_pairs(self):
for name in os.listdir(self.data_dir):
if name == ".DS_Store":
continue
a = []
for file in os.listdir(self.data_dir + '/' + name):
if file == ".DS_Store":
continue
a.append(name + '/' + file)
generatePairs.write_similar(a)
def generate_non_pairs(self):
folder_list = []
for folder in os.listdir(self.data_dir):
folder_list.append(folder)
folder_list.sort(reverse=True)
# print(folder_list)
i = 0
a = []
for dir in os.listdir(self.data_dir):
if dir == ".DS_Store":
continue
for file in os.listdir(self.data_dir + dir):
if file == ".DS_Store":
continue
a.append(dir + '/' + file)
# print(a)
b = []
for dir in os.listdir(self.data_dir):
if dir == ".DS_Store":
continue
for file in os.listdir(self.data_dir + folder_list[i]):
if file == ".DS_Store":
continue
b.append(folder_list[i] + '/' + file)
# print(b)
i = i + 1
generatePairs.write_different(a, b)
if __name__ == '__main__':
# data_dir = "/home/ti/Downloads/DATASETS/out_data_crop/"
# pairs_filepath = "/home/ti/Downloads/insightface/src/data/pairs.txt"
# alternative_lst = "/home/ti/Downloads/insightface/src/data/crop.lst"
# test_txt = "/home/ti/Downloads/DATASETS/out_data_crop/test.txt"
# img_ext = ".png"
# arguments to pass in command line
parser = argparse.ArgumentParser(description='Rename images in the folder according to LFW format: Name_Surname_0001.jpg, Name_Surname_0002.jpg, etc.')
parser.add_argument('--dataset-dir', default='', help='Full path to the directory with peeople and their names, folder should denote the Name_Surname of the person')
parser.add_argument('--list-file', default='', help='Full path to the directory with peeople and their names, folder should denote the Name_Surname of the person')
parser.add_argument('--img-ext', default='', help='Full path to the directory with peeople and their names, folder should denote the Name_Surname of the person')
# reading the passed arguments
args = parser.parse_args()
data_dir = args.dataset_dir
lst = args.list_file
img_ext = args.img_ext
# generatePairs = PairGenerator(data_dir, pairs_filepath, img_ext)
# generatePairs.write_item_label()
# generatePairs = PairGenerator(data_dir, pairs_filepath, img_ext)
generatePairs = PairGenerator(data_dir, lst, img_ext)
generatePairs.write_item_label_abc() # looping through our dataset and creating 1 ITEM_PATH LABEL lst file
# generatePairs.generate_pairs() # to use, please uncomment this line
# generatePairs.generate_non_pairs() # to use, please uncomment this line
# generatePairs = PairGenerator(dataset_dir, test_txt, img_ext)
# generatePairs.split_to_10()
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
附录3 CFP数据集生成pair文件
import mxnet as mx
from mxnet import ndarray as nd
import argparse
import pickle
import sys
import os
parser = argparse.ArgumentParser(description='Package LFW images')
# general
parser.add_argument('--data-dir', default='', help='')
parser.add_argument('--image-size', type=str, default='112,112', help='')
parser.add_argument('--output', default='', help='path to save.')
args = parser.parse_args()
data_dir = args.data_dir
image_size = [int(x) for x in args.image_size.split(',')]
pairs_end = []
def get_paths():
pairs = []
prefix = os.path.join(data_dir,'Protocol/')
#prefix = "/Split/"
prefix_F = os.path.join(prefix, "Pair_list_F.txt")
pairs_F = []
prefix_P = os.path.join(prefix,"Pair_list_P.txt")
pairs_P = []
pairs_end = []
issame_list = []
#读pairlist文件
with open(prefix_F, 'r') as f:
for line in f.readlines()[0:]:
pair = line.strip().split()
pairs_F.append(pair[1])
print(len(pairs_F))
with open(prefix_P, 'r') as f:
for line in f.readlines()[0:]:
pair = line.strip().split()
pairs_P.append(pair[1])
print(len(pairs_P))
#读pair文件
prefix = os.path.join(data_dir,"Protocol/Split/FP")
folders_1 = os.listdir(prefix)
for folder in folders_1:
sublist = []
same_list = []
pairtxt = os.listdir(os.path.join(prefix, folder))
for pair in pairtxt:
img_root_path = os.path.join(prefix, folder, pair)
with open(img_root_path, 'r') as f:
for line in f.readlines()[0:]:
#print(line)
pair1 = line.strip().split(',')
#print(pair)
pairs_end += (os.path.join(data_dir,'Protocol/',pairs_F[int(pair1[0])-1]),os.path.join(data_dir,'Protocol/',pairs_P[int(pair1[1])-1]))
#print(pair)
if pair == 'same.txt':
#print('ok!')
issame_list.append(True)
else:
issame_list.append(False)
return pairs_end,issame_list
lfw_paths, issame_list = get_paths()
lfw_bins = []
print(len(lfw_paths))
print(lfw_paths[0])
print(lfw_paths[1])
print(issame_list[0])
print(issame_list[1])
#lfw_data = nd.empty((len(lfw_paths), 3, image_size[0], image_size[1]))
print(len(issame_list))
i = 0
for path in lfw_paths:
with open(path, 'rb') as fin:
_bin = fin.read()
lfw_bins.append(_bin)
#img = mx.image.imdecode(_bin)
#img = nd.transpose(img, axes=(2, 0, 1))
#lfw_data[i][:] = img
i+=1
if i%1000==0:
print('loading lfw', i)
with open(args.output, 'wb') as f:
pickle.dump((lfw_bins, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
1
有问题请留言，最近一两周在做这个可以更新帖子及问题
Next：
insightface测试验证集效果全过程
https://blog.csdn.net/CLOUD_J/article/details/98882718
4145
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
