Lightweight Face Recognition

• Framework : Python 3.7 + , Pytorch 1.4.0, FastAPI 0.1.0

MobileFaceNet

The implementation of popular face recognition algorithms in pytorch framework, including arcface.
In this repository, we use a pre-trained MobliefaceNet model.

Requirements

for Face recognizer

PyQt5 pillow requests opencv-python bcolz matplotlib numpy pandas scikit-learn tensorboardX tqdm

for Rest API

fastapi uvicorn email_validator aiofiles sqlalchemy mysql-connector-python pyjwt passlib[bcrypt] python-multipart

Apis

Identify

POST /api/identify
{
   group_id: str
   threshold: float = None
   img : str # base64 bytes string
}

Adding a person

POST /api/person
{
   group_id : str
   person_id : str
   person_name : str = None
   img : str # base64 bytes string
}

References

wujiyang/Face_Pytorch
MuggleWang/CosFace_pytorch
foamliu/MobileFaceNet-PyTorch
Xiaoccer/MobileFaceNet_Pytorch
TreB1eN/InsightFace_Pytorch
deepinsight/insightface
KaiyangZhou/pytorch-center-loss
tengshaofeng/ResidualAttentionNetwork-pytorch

Install

1. download and unzip or [git clone]
   download this project and unzip
   move the directory
   cd [project]

2. install anaconda 3.x

3. create virtual environment and activate
   conda create -n [project name] python=3.7
conda activate [project name]

4. download pre-trained model
   download latest model_mobilefacenet.pth to the [project] directory

5. (additional) for GPU version
   install CUDA Toolkit
   install cuDNN

6. install Pytorch and Cuda toolkit

   • GPU (nvidia product only)
     conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
     (need to check your cuda version)
   • CPU
     conda install pytorch torchvision cpuonly -c pytorch

7. (additional) for windows > before install requirements install "Microsoft Build Tools 2015 Update 3" for bcolz package
   (visualcppbuildtools_full.exe)

8. install requirement packages
   pip install -r requirements.txt
   (could not install some packages by conda install)

Run server

cd [project] directory
python server.py

Test in web browser

• open chrome
  http://127.0.0.1:8000/docs