I2I-GANs

Common generative adversarial networks (GANs) implemented in TensorFlow 2.4.1. The GANs are suitable for image-to-image translation tasks.

The repository was published as a part of the master's thesis (Generative Adversarial Networks Applied for Privacy Preservation in Biometric-Based Authentication and Identification). Preliminary results were presented at http://excel.fit.vutbr.cz/submissions/2021/031/31.pdf.

The following architectures are implemented:

• TraVeLGAN (https://github.com/KrishnaswamyLab/travelgan)
• DiscoGAN (https://github.com/SKTBrain/DiscoGAN)
• GcGAN (https://github.com/hufu6371/GcGAN)

Setup

1. Clone this repository:

   git clone https://github.com/lubosmj/I2I-GANs && cd I2I-GANs
   

2. Create a new virtual environment:

   python3 -m venv venv
   source source venv/bin/activate
   

3. Install the packages:

   python3 setup.py install
   

4. Use the installed modules in your application:

   from i2i_gans import TraVeLGAN
   
   travelgan = TraVeLGAN(...)
   travelgan.compile()
   travelgan.load_weights(...)
   
   fake_images = travelgan.generator(...)

Running the Examples

1. Train a new TraVeLGAN model:

   python3 -m examples.travelgan_trainer train --domain_A "path/to/dataset/A/*.png" --domain_B "path/to/dataset/B/*.png" --dataset_size 5000 --batch_size=16 --checkpoints_freq 10 --parallel --samples_freq 10 --samples_dir samples --checkpoints_dir checkpoints --augment random_flip_left_right --epochs 250
   

2. Train a new DiscoGAN model:

   python3 -m examples.discogan_trainer train --domain_A "path/to/dataset/A/*.png" --domain_B "path/to/dataset/B/*.png" --dataset_size 5000 --batch_size=200 --checkpoints_freq 10 --parallel --samples_freq 10 --samples_dir samples --checkpoints_dir checkpoints --augment random_flip_left_right --epochs 200
   

3. Train a new GcGAN model:

   python3 -m examples.gcgan_trainer train --domain_A "path/to/dataset/A/*.png" --domain_B "path/to/dataset/B/*.png" --dataset_size 5000 --batch_size=12 --checkpoints_freq 10 --parallel --samples_freq 10 --samples_dir samples --checkpoints_dir checkpoints --augment random_flip_left_right --epochs 200
   

Generated Images

TraVeLGAN

The GAN was trained for 250 epochs with Adam optimizer (learning rate: 0.0002, batch size: 16, dataset size: 8,000).

• Datasets:
  • Augmented images from CelebA
  • flowers102

DiscoGAN

The GAN was trained for 200 epochs with the same hyper-parameters as recommended in the original paper (dataset size: 20,000). Additionally, one convolution layer with 100 filters was inserted into the generators.

• Datasets:
  • CelebA
  • UT Zappos50K