Face-Recognition-with-InsightFace
MTCNN
InsightFace

pip install keras
pip install tensorflow==2.2

datasets\train里的目录名中的空格’ ‘改成’_’
使用datasets\train里的图片生成train.lst
修改src\data\dir2lst.py

import sys
import os
sys.path.append('../common')
import face_image

input_dir = sys.argv[1]

dataset = face_image.get_dataset_common(input_dir, 2)
output_filename = os.path.join(input_dir, 'train.lst')
with open(output_filename, "a") as text_file:
    for item in dataset:
      oline = "%d\t%s\t%d\n" % (1, item.image_path, int(item.classname))
      text_file.write(oline)

python dir2lst.py F:\insightface-master0\datasets\train

修改src\data\face2rec2.py

#try:
    #import multiprocessing #Windows下
#except ImportError:


s = mx.recordio.pack(header, b'')#加上b

python face2rec2.py F:\insightface-master0\datasets\train

使用generate_image_pairs.py
并修改

    if len(same_list) > 10: #and len(same_list) < 13
        for j in range(0, 10, 2): #len(same_list)

python generate_image_pairs.py --data-dir F:\insightface-master0\datasets\train --outputtxt F:\insightface-master0\datasets\train\train.txt --num-samepairs 100

修改lfw2pack.py

python lfw2pack2.py --data-dir F:\insightface-master0\datasets\train --output F:\insightface-master0\datasets\train\train.bin --num-samepairs 100

生成的文件放到recognition\datasets\train里
新建文件property
内容为类别数,112,112

recognition\ArcFace
复制sample_config.py为config.py
修改config.py

dataset.emore.dataset = 'emore'
dataset.emore.dataset_path = '../datasets/train'
dataset.emore.num_classes = #类别数#
dataset.emore.image_shape = (112, 112, 3)
dataset.emore.val_targets = ['train']


default.end_epoch = 100 
default.per_batch_size = 32 #128# 显卡垃圾

修改verification.py

    #val = float(true_accept) / float(n_same)
    #far = float(false_accept) / float(n_diff)
    #改成
    if n_same == 0:
        val = 1
    else:
        val = float(true_accept) / float(n_same)
    if n_diff == 0:
        far = 0
    else:
        far = float(false_accept) / float(n_diff)

修改train.py

#_rescale = 1.0 / args.ctx_num
_rescale = 0.03125

然后训练并验证

set CUDA_VISIBLE_DEVICES='0,' 
python -u train.py --network r100 --loss arcface --dataset emore

修改src\train_softmax.py

    #print(his.history['accuracy'])

    history['acc'] += his.history['accuracy']
    history['val_acc'] += his.history['val_accuracy']

参考RetinaFace\test.py修改src\recognizer_image.py

from retinaface import RetinaFace
...
detector = RetinaFace('../RetinaFace/model/R50', 0, gpuid, 'net3')
...
faces, landmarks = detector.detect(img, thresh, scales=scales, do_flip=flip)
...

python faces_embedding.py --dataset F:\insightface-master0\datasets\train 
python train_softmax.py
python recognizer_image.py --image-in ../datasets/test/005.jpg

其它
下载lfw-deepfunneled
使用src\align\align_lfw.py 生成对齐后的人脸
或参考编译RetinaFace及使用

_paths = fimage.image_path.split('\\')#Windows下'/'改成'\\'

py