Face Aging with Generative Adversarial Network Project Report

This project was developed for 01URXOV - Machine Learning in Applications (spring semester 2022) of Politecnico di Torino.

Authors: Mattia Dutto, Julian Neubert, Simone Alberto Peirone.

Contributions

The starting point is the repo FaceAgin-by-cycleGAN that it used for the baseline approach. The starting point of AE-StyleGAN is taken from repo

Our contribution mainly focuses on the use of CelebA dataset on the two approaches and on the Cycle-StyleGAN, that is based on the previously cited AE-StyleGAN.

Scripts details:

image_folder_cyclegan.py: it allows to use CelebA dataset to the FaceAging-by-cycleGAN architecture. There are basically custom files of the one present in the repo.
split-celaba.py: it allows to create the test, validation and train split of the selected dataset.
sample-celeba.py: it allows use to split the dataset partitions (train, validation, test) based on a attribute by creating 2 sub-folders (in this case the young and old one). This script does not copy the source images to the destination file, it will just create symlinks.
train_cyclegan.py: is the train.py customized for our task.
envinroments_cyclegan.yml: is for install all the dependencies of Face-Aging-by-cycleGAN.

Scripts folder

Inside this folder you can find the examples of the sbatch scripts for the different networks on the cluster of Politecnico of Turin.

run_cycle_gan.sbatch: Cycle-GAN training
run_ae_stylegan.sbatch: AE-StyleGAN training
run_ae_stylegan_ablation.sbatch: hyper-parameter search on AE-StyleGAN
run_ae_cycle_stylegan.sbatch: Cycle-StyleGAN training.

List of task to do before running any experiment.

Clone this repo and clone inside the FaceAging-by-cycleGAN repo

Create the environment: with torchvision and the requirement present on FaceAging-by-cycleGAN/requirements.txt

conda env create -n fa -f FaceAging-by-cycleGAN/environment.yml
source activate fa
pip install torchvision
pip install -r FaceAging-by-cycleGAN/requirements.txt
conda install cython

CelebA dataset:
- Download the dataset from here.
- Unzip the dataset
- Run the split-celeba.py script
- Run the sample-celeba.py script
```
unzip -qq img_align_celeba.zip

python split-celeba.py
python sample-celeba.py
```

Move pre-trained weights to the correct place and rename them (the 9_wiki_fine_tune_male pre-trained model was used and the latest part must be changed with 200)

# note: You need to be on the home directory of FaceAging-by-cycleGAN repo.

mv trained_model/9_wiki_fine_tune_male/* checkpoints/9_wiki_fine_tune_male/

mv latest_net_D_A.pth 200_net_D_A.pth
mv latest_net_D_B.pth 200_net_D_B.pth
mv latest_net_G_A.pth 200_net_G_A.pth
mv latest_net_G_B.pth 200_net_G_B.pth

Move the custom files to the correct directory.

# note: position yourself on the home directory of this repo.
mv data_loader_cyclegan.py ./FaceAging-by-cycleGAN/data/__init__.py

Example of test command for Cycle-GAN

This command allows to generate one of the results that you can see on the repo.

python test.py --dataroot ../celeba \
               --name 9_wiki_fine_tune_male/05_epochs \
               --model cycle_gan  \
               --epoch end \
               --gpu_ids -1 \
               --num_test 8