search

garynlfd / KFC

KFC: Kinship Verification with Fair Contrastive loss and Multi-Task Learning
MIT License
6 stars 1 forks source link

readme

KFC 🍗 (BMVC 2023)

KFC: Kinship Verification with Fair Contrastive loss and Multi-Task Learning

Jia Luo Peng,  Keng Wei Chang,  Shang-Hong Lai  National Tsing Hua University, Taiwan [![arXiv](https://img.shields.io/badge/arXiv-<2309.10641>-.svg)](https://arxiv.org/abs/2309.10641)

News

2023-08: Accepted to BMVC 2023

Introduction

KFC: Kinship Verification with Fair Contrastive Loss and Multi-Task Learning aims to solve the kinship verification task while improving racial fairness. We utilize an attention module in our model and employ multi-task learning to enhance accuracy in kinship verification. Additionally, we have developed a new fair contrastive loss function and use gradient reversal to decrease the standard deviation among the four races (African, Asian, Caucasian, Indian).

Furthermore, we combine six kinship datasets (see below) to create a large kinship dataset. Moreover, we have annotated every identity in our dataset to include racial information.

Requirements

Installation

  1. Clone this project and create virtual environment 
    $ git clone https://github.com/garynlfd/KFC.git
$ conda create --name KFC python=3.8
$ conda create activate KFC
  2. Install requirements 
    $ pip install -r requirements.txt

    Datasets

  3. Please download these datasets from this URL
  4. These are the datasets I used. I downloaded them and found that some of the images are wrong person, so I did data cleaning. So please use the above URL to run the code without error.
  5. Please place these datasets in the same folder as train.py, find.py and test.py: 
    ./KFC
├── train.py
├── find.py
├── test.py
├── ...(other files)
├── Cornell_Kin/
├── UB_KinFace/
├── KinFaceW-I/
├── KinFaceW-II/
├── Family101_150x120/
├── Train(from FIW)/
├── Validation(from FIW)/
└── Test(from FIW)/

    Command

    Training

    batch_size: default 25
    sample: The folder in which the data should be placed, corresponding to data files folder
    save_path: THe folder in which the ckpt will be saved, corresponding to log files folder
    epochs: default 100
    beta: temperature parameters default 0.08
    log_path: name the log file
    gpu: choose which gpu you want to use 

    $ python train.py --batch_size 25 \
--sample ./data_files \
--save_path ./log_files \
--epochs 100 --beta 0.08 \
--log_path log_files/{file_name}.txt \
--gpu 1

    Finding

    sample: The folder in which the data should be placed, corresponding to data files folder
    save_path: THe folder in which the ckpt will be saved, corresponding to log files folder
    batch_size: default 50
    log_path: name the log file
    gpu: choose which gpu you want to use 

    $ python find.py --sample ./data_files \
--save_path ./log_files \
--batch_size 50 \
--log_path log_files/{file_name}.txt \
--gpu 1

Testing

sample: The folder in which the data should be placed, corresponding to data files folder
save_path: THe folder in which the ckpt will be saved, corresponding to log files folder
threshold: the output value from find.py
batch_size: default 50
log_path: name the log file
gpu: choose which gpu you want to use 

$ python test.py --sample ./sample0 \
--save_path ./log_files \
--threshold {output value from find.py} \
--batch_size 50 \
--log_path log_files/{file_name}.txt \
--gpu 1

Acknowledgements

Our implementation uses code from the following repositories: