Orthogonal Jacobian Regularization for Unsupervised Disentanglement in Image Generation

Home | PyTorch BigGAN Discovery | TensorFlow ProGAN Regularization | PyTorch Simple GAN Experiments | Paper

Simple Complex Left Complex Left Complex Left Complex Left

This repo contains code for our OroJaR Regularization that encourages disentanglement in neural networks. It efficiently optimizes the Jacobian vectors of your neural network with repect to each input dimension to be orthogonal, leading to disentanglement results.

This repo contains the following:

Adding the OroJaR to Your Code

We provide portable implementations of the OroJaR that you can easily add to your projects.

Adding the OroJaR to your own code is very simple:

from orojar_pytorch import orojar

net = MyNeuralNet()
input = sample_input()
loss = orojar(G=net, z=input)

Getting Started

This section and below are only needed if you want to visualize/evaluate/train with our code and models. For using the OroJaR in your own code, you can copy one of the files mentioned in the above section.

Both the TensorFlow and PyTorch codebases are tested with Linux on NVIDIA GPUs. You need at least Python 3.6. To get started, download this repo:

git clone https://github.com/csyxwei/OroJaR.git
cd OroJaR

Then, set-up your environment. You can use the environment.yml file to set-up a Conda environment:

conda env create -f environment.yml
conda activate orojar

If you opt to use your environment, we recommend using TensorFlow 1.14.0 and PyTorch >= 1.6.0. Now you’re all set-up.

TensorFlow ProgressiveGAN Regularization Experiments

PyTorch BigGAN Direction Discovery Experiments

Other Experiments with Simple GAN


If our code aided your research, please cite our paper:

  title={Orthogonal Jacobian Regularization for Unsupervised Disentanglement in Image Generation},
  author={Wei, Yuxiang and Shi, Yupeng and Liu, Xiao and Ji, Zhilong and Gao, Yuan and Wu, Zhongqin and Zuo, Wangmeng},
  booktitle={Proceedings of International Conference on Computer Vision (ICCV)},