This is the official repository of the paper:

AdaDistill: Adaptive Knowledge Distillation for Deep Face Recognition

Accepted at ECCV 2024 (main conference)

Arxiv

Installation

• The code is running using Pytorch 1.7
• Download the requirements from the file requirement.txt
• Download the processed MS1MV2 from the MS1MV2, unzip it and place it inside the data folder
• The code is originally designed to run on 4 GPUs which can be changed from running scripts, run_standalone.sh

Teacher and standalone student training

• Set the config.network parameter in the config/config.py to iresent50 or mobilefacenet
• Set the output folder where the model should be saved
• The teacher and standalone student can be trained by running ./run_standalone.sh

ArcDistill and CosDistill training

• Set the config.network parameter in the config/config.py to mobilefacenet
• Set the output folder where the model should be saved in the config/config.py
• Set the path to the teacher header and backbone from previous training in the config/config.py , config.pretrained_teacher_path and config.pretrained_teacher_header_path
• Set the penalty loss to ArcFace or to CosFace in the config/config.py
• ArcDistill and CosDistill can be trained by running ./run_AMLDistill.sh

AdaDistill training

• Set the config.network parameter in the config/config.py to mobilefacenet
• Set the parameter config.adaptive_alpha in the config/config.py to False
• Set the output folder where the model should be saved in the config/config.py
• Set the path to the teacher backbone in the config/config.py, config.pretrained_teacher_path
• Set the penalty loss to ArcFace or to CosFace in the config/config.py
• AdaArcDistill and AdaCosDistill can be trained by running ./run_AdaDistill.sh

AdaDistill training with weighted alpha

• Set the config.network parameter in the config/config.py to mobilefacenet
• Set the parameter config.adaptive_alpha in the config/config.py to True
• Set the output folder where the model should be saved in the config/config.py
• Set the path to the teacher backbone in the config/config.py, config.pretrained_teacher_path
• Set the penalty loss to ArcFace or to CosFace in the config/config.py
• AdaArcDistill and AdaCosDistill can be trained by running ./run_AdaDistill.sh

A trained MobileFaceNet model with AdaArcDsitll is provided under output/AdaDistill/MFN_AdaArcDistill_backbone.pth

If you use any of the code provided in this repository, please cite the following paper:

Citation

@InProceedings{Boutros_2024_ECCV,
    author    = {Fadi Boutros, Vitomir Štruc, Naser Damer},
    title     = {AdaDistill: Adaptive Knowledge Distillation for Deep Face Recognition},
    booktitle = {Computer Vision - {ECCV} 2024 -18th European Conference on Computer Vision, Milano, Italy, September 29- 4 October, 2024 },
    month     = {October},
    year      = {2024},
    pages     = {}
}

License

This project is licensed under the terms of the Attribution-NonCommercial-ShareAlike 4.0 
International (CC BY-NC-SA 4.0) license. 
Copyright (c) 2024 Fraunhofer Institute for Computer Graphics Research IGD Darmstadt