Interpretable Face Aging: Enhancing Conditional Adversarial Autoencoders with LIME Explanations

This repository was created with the aim of incorporating an Explanatory System in the Conditional Adversarial Autoencoder (CAAE). The main source of inspiration was the paper Age Progression/Regressionby Conditional Adversarial Autoencoder proposed by (Zhang, Song, et al.). Moreover, the paper xAI-GAN: Enhancing Generative Adversarial Networks via Explainable AI Systems proposed by (Nagisetyy, Graves, et al.) formed the base of adding xAI methods in CAAE. Finally, thanks to Mattan Serry and the AgeProgression work, we developed our code in PyTorch.

Our Paper

This repo contains the codes of the proposed Interpretable Conditional Adversarial Autoencoder (ICAAE) described in the Interpretable Face Aging: Enhancing Conditional Adversarial Autoencoders with LIME Explanations paper. The paper is published in the proceedings of the 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2024).

Dataset

The dataset we used for training was the CACD + UTKFace which consists of 21267 images in 7 age classes. FGNET with 1002 images was used for the test purpose. The data folder can be downloaded through Google Drive.

Prerequisites

1. Python 3.7
2. Pytorch 1.2.0
3. Captum 0.4.0
4. Lime 0.2.0.1

Implementation

The control and change of the Explainable Artificial Intelligence system is achieved through the net.teachSplit() in mainCAAEsplit_ver2_lime.py script. If it is a need to train the Original CAAE, without the addition of the xAI system, explainable is set to False. The xAI system is activated by setting explainable to True. LIME can be added by setting the explanation_type to 'lime'.

Virtual Environment

conda create -n icaae python=3.7 anaconda

conda activate icaae

pip install -r requirements.txt

Training

python mainCAAEsplit_ver2_lime.py --mode train --epochs 140 --input data/CACD_UTKFace --output checkpoints

Testing

python mainCAAEsplit_ver2_lime.py --mode test --load checkpoints/epoch140 --input data/FGNET --output results/checkpoints/epoch140 --age <0 to 6> --gender <0 or 1>

Directory Tree

LIME-CAAE   
│   consts.py  
│   mainCAAEsplit_ver2_lime.py  
│   modelSplit_v2_lime.py   
│   README.md
│   my_lime_image.py
|   requirements.txt
│   utils.py
|   utils_xai.py
└───data
    └───CACD_UTKFace
    └───FGNET

Citation

@INPROCEEDINGS{10447128,
 author={Korgialas, Christos and Pantraki, Evangelia and Kotropoulos, Constantine},
 booktitle={ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
 title={Interpretable Face Aging: Enhancing Conditional Adversarial Autoencoders with Lime Explanations}, 
 year={2024},
 volume={},
 number={},
 pages={5260-5264},
 doi={10.1109/ICASSP48485.2024.10447128}}

Authors

Feel free to send me a message for any issue.

Christos Korgialas (ckorgial@csd.auth.gr), Evangelia Pantraki (epantrak@csd.auth.gr), and Constantine Kotropoulos (costas@csd.auth.gr)