News:1 paper accepted by CVPR2018

Contents

1. 1. 1. Installation
2. 2. 2. Download datasets
3. 3. 3. Test on saved models
4. 4. 4. Train from scratch
5. 5. 5. Experiment results
6. 6. Citation

This repo is the official open source of Face Aging with Identity-Preserved Conditional Generative Adversarial Networks, CVPR 2018 by Zongwei Wang, Xu Tang, Weixin Luo and Shenghua Gao.
It is implemented in tensorflow. Please follow the instructions to run the code.

1. Installation

• Install 3rd-package dependencies of python (listed in requirements.txt)

  tensorflow-gpu==1.4.1
  scipy==1.0.0
  opencv-python==3.3.0.10
  numpy==1.11.0
  Pillow==5.1.0

pip install -r requirements.txt

• Other libraries

  CUDA 8.0
  Cudnn 6.0

2. Download datasets

We use the Cross-Age Celebrity Dataset for training and Evaluation. More details about this dataset, please refer to (http://bcsiriuschen.github.io/CARC/). After face detection, aligning and center cropping, we split
images into 5 age groups: 11-20, 21-30, 31-40, 41-50 and 50+.

3. Test on saved models

Download the trained face aging model(https://1drv.ms/u/s!AlUWwwOcwDWobCqmuFyKGIt4qaA) and place models files in checkpoints/0_conv5_lsgan_transfer_g75_0.5f-4_a30.

• Test images are in images/test, and some training images that belong to 11-20 age group are in images/train.

  * Running the sript to get aged faces
      python test.py

4. Train from scratch

• Firstly, download the pre-trained alexnet model(https://1drv.ms/u/s!AlUWwwOcwDWobkptownyu5fjlfU) and age classfication model(https://1drv.ms/f/s!AlUWwwOcwDWocX-Z0IJft_VbcoQ). Then unzip these files and place model files in checkpoints/pre_trained.

  python age_lsgan_transfer.py \
    --gan_loss_weight=75 \
    --fea_loss_weight=0.5e-4 \
    --age_loss_weight=30 \
    --fea_layer_name=conv5 \
    --checkpoint_dir=./checkpoints/age/0_conv5_lsgan_transfer_g75_0.5f-4_a30 \
    --sample_dir=age/0_conv5_lsgan_transfer_g75_0.5f-4_a30

sh age_lsgan_transfer.py

• You can change the hyperparameters to adapt to your own dataset.

5. Experiment results

• The aging effect of different methods. Within each dot box, the first row is our result.
• The aging effect of different age classification loss weights.
• The aging effect of different feature layer.

Citation

If you find this useful, please cite our work as follows:

@INPROCEEDINGS{wang2018face_aging, 
    author={Z. Wang and X. Tang, W. Luo and S. Gao}, 
    booktitle={2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)}, 
    title={Face Aging with Identity-Preserved Conditional Generative Adversarial Networks}, 
    year={2018}
}