Multi-task Face Recognition Framework

Introduction

This repository is a multi-task face recognition framework, built on top of the PyTorch implementation of ArcFace.
It is for our IJCNN'20 paper Neighborhood-Aware Attention Network for Semi-supervised Face Recognition.
You may refer to the repository NAAN for the fully semi-supervised implementation of our paper.

Requirments

• python >= 3.5
• pytorch >= 1.0.0
• numpy
• tensorboardX

Data Preparation

• Download the full MS-Celeb-1M realeased by ArcFace from baidu or dropbox, and move them to the folder faces_emore.
• Download the splitted image list produced by learn-to-cluster from GoogleDrive or OneDrive, and move them to the folder lists.
• Re-arrange the dataset using preprocess.py. The folder structure of emore is the same as:

  emore
   ├── trainset
   ├── testset
      ├── split_1
      ├── split_2
      ├── split_3
      ├── split_4
      ├── split_5
      ├── split_6
      ├── split_7
      ├── split_8
      ├── split_9
  

Training

sh train_multi.sh

Modify the param path in train_multi.sh to the directory of the generated pseudo-label file split{}_labels.txt.