anne41326 / Insightface_arcface_pytorch

Insightface_arcface_pytorch

This repo just recorded the project I have made last few months. The project uses ArcFace, and based on Arcface, I provide dataset_prepare.py for own dataset prepareing, and also provide create_pairs.py to create pairs.txt for testing.

Requirements

numpy
torch
PIL
torchvision
tqdm
splitfolders

Dataset prepare

1. First put your photos in folders respectively. The structure just like the following

A(folder)
    12443.jpg
    423435.jpg
    ...
B (folder)
    64344.jpg
    34353.jpg
    ...

2. Run python dataset.prepare.py

3. After running, the structure may become the following

照片資料夾-
    資料夾A
            - A_00.jpg
            - A_01.jpg
    資料夾B
            - B_00.jpg
            - B_01.jpg
            ...
            
    train(folder，每個folder中至少放一張沒戴口照照片)
            -資料夾A
            - A_00.jpg
            - A_09.jpg(照片數量是原本的7成)
            ...
    val(folder)
            -資料夾A
            - A_02.jpg
            - A_04.jpg(照片數量是原本的3成)
            ...
# If you want model could recognize pepole with mask, you have to put at least one no masked photo in each folder in train folder manually

As seen, train folder(70% photos of original photos) and val folder(30% photos of original photos) are created, and there are photos in each folder.

Train

1. Run python train.py
2. After training, there may be a 23.pth file in checkpoints folder

Test

1. First create pairs.txt for testing. Run python create_pairs.py
2. Make sure pairs.txt exists and is normal like this(1 stands for the same person, 0 stands for other person)

04\04_5.jpg 04\04_1.jpg 1
40\40_3.jpg 40\40_7.jpg 1
05\05_1.jpg 05\05_0.jpg 1
27\27_2.jpg 27\27_3.jpg 1
34\34_7.jpg 30\30_3.jpg 0
49\49_2.jpg 13\13_8.jpg 0
36\36_4.jpg 60\60_0.jpg 0
...

3. Run test_model.py

4. Finally, you will get accuracy,threshold

Contribution and Restriction

1. arcface-pytorch

2. Build-Your-Own-Face-Model : there is a mistake on comment ( actually is not distance ,but similarity)

• for more , please check https://blog.csdn.net/weixin_43977640/article/details/115579153 https://www.cnblogs.com/suanec/p/9121092.html

3. split-folders