This repository provides PyTorch implementations for RankCGAN paper published in BMVC 2018.

This code is based on DCGAN implementation pytorch-DCGAN.

The model architecture comprises three modules, the discriminator, generator and ranker. As shown in the following Figure:

• UT-Zap50K
• Pubfig
• OSR

• UT-Zap50K
• Pubfig
• OSR

In this code, only UT-Zap50K dataset is used. For other datasets, please edit dataloader.py accordingly.

• Python 3.5.2
• Pytorch 0.4.0
• NVIDIA GPU + CUDA CuDNN

• Install PyTorch 0.4+ and torchvision from http://pytorch.org.
• Clone this repository.

git clone https://github.com/saquil/RankCGAN
cd RankCGAN

The code provides two implementations. RankCGAN.py demonstrates one attribute conditional model and RankCGAN_2D.py demonstrates multiple attributes case with two attributes conditional model.

• Train a model on single attribute:

python3 main.py --gan_type=RankCGAN

• Train a model on multiple attributes:

python3 main.py --gan_type=RankCGAN_2D

• 1D images generation:

Generated shoe (top), face (middle), and scene (bottom) images associated with their ranking scores using “sporty”, “masculine” and “natural” attributes respectively.

• 2D images generation:

Example of two-attributes interpolation on shoe and face images using (“sporty",“black") and (“masculine",“smiling") attributes.

If you use this code for your research, please cite our paper.

@inproceedings{saquil2018ranking,
  title={Ranking CGANs: Subjective Control over Semantic Image Attributes},
  author={Saquil, Yassir and Kim, Kwang In and Hall, Peter}
  booktitle={British Machine Vision Conference (BMVC)},
  year={2018}
}

• You can find our BMVC 2018 poster here
• You can find our Supplementary Material here