Weakly Supervised Disentanglement by Pairwise Similarities

This repository contains the PyTorch source code for the paper Weakly Supervised Disentanglement by Pairwise Similarities by Junxiang Chen and Kayhan Batmanghelich.

Model

In this paper, we propose a setting where the user introduces weaksupervision by providing similarities between instances (denoted by $y_{ij}$) based on a factor to be disentangled. The similarity is providedas either a binary (yes/no) or real-valued label describingwhether a pair of instances are similar or not. We propose a new method for weakly supervised disentanglement of latent variables within the framework of Variational Autoencoder.

Environment

To prepare for the environment for running our code, run

conda env create -f VAE_pairwise.yml

Experiments

We include an example of training script in test.ipynb.

MNIST with binary pairwise labels

We train the model with binary pairwise labels for the MNIST dataset. The embedding and generated results are shown below:

MNIST with real-valued pairwise labels

We also train the model with real-valued pairwise labels for the MNIST dataset. This setting is for illustration purposes only, but might not be useful in solving real-world problems. The embedding and generated results are shown below:

Citation

@article{chen2019weakly,
  title={Weakly Supervised Disentanglement by Pairwise Similarities},
  author={Chen, Junxiang and Batmanghelich, Kayhan},
  journal={arXiv preprint arXiv:1906.01044},
  year={2019}
}

Acknowledgments

This work was partially supported by NIH Award Number 1R01HL141813-01, NSF 1839332 Tripod+X, and SAP SE. We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan X Pascal GPU used for this research. We were also grateful for the computational resources provided by Pittsburgh SuperComputing grant number TG-ASC170024.