Lr-LiVAE (CVPR2019)

Tensorflow implementation of Disentangling Latent Space for VAE by Label Relevant/Irrelevant Dimensions.

The network architecture

Exemplar Results

• Toy experiment

• Generated images

• Face synthesis

Preparation

• Prerequisites
  • Tensorflow
  • Python 2.x with matplotlib, numpy and scipy
• Dataset
  • FaceScrub
    • Images should be placed in ./Datas/facescrub_aligned/
• pre-trained model
  • Download the trained model from Baidu Cloud(dfna), and unzip the files to ./models/

Quick Start

Exemplar commands are listed here for a quick start.

Training

• To train on facescrub dataset with size of 64 X 64

  python Lr-LiVAE.py --mode training

Testing

• Example of generating random samples

  python Lr-LiVAE.py --mode generation

• Example of image inpainting

  python Lr-LiVAE.py --mode inpainting

• Example of face synthesis using z_s and z_u from different images

  python Lr-LiVAE-GAN.py --mode exchanging

Citation

If this work is useful for your research, please consider citing:

@inproceedings{zheng2019disentangling,
  title={Disentangling Latent Space for VAE by Label Relevant/Irrelevant Dimensions},
  author={Zheng, Zhilin and Sun, Li},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={12192--12201},
  year={2019}
}