Independent Vector VAE

PyTorch implementation for paper

• "Semi-supervised Disentanglement with Independent Vector Variational Autoencoders," arXiv preprint arXiv:2003.06581 (2020)

Tested Environment

• {python 2.7, pytorch 0.4.1} OR
• {python 3.6, pytorch 1.1.0}

Dataset

• Use the following script (~570MB storage needed):

bash ./download_data.sh

• Alternative: download the data via this GoogleDrive link → put them in data/
• Note: the images were resized and divided into training, validation, and test sets as follows.
  • Fashion-MNIST, MNIST: 32×32, # (train, valid, test) imgs = (50K, 10K, 10K)
  • dSprites: 64×64, # (train, valid, test) imgs = (614K, 61K, 61K)

Code

• For Fashion-MNIST and MNIST, run the following scripts. Both use the separate TC setup.

bash ./exp_fashMni.sh
bash ./exp_mnist.sh

• For dSprites under the separate TC setup, run:

bash ./exp_shapes_sepaTc.sh

• For dSprites under the collective TC setup, run:

bash ./exp_shapes_collecTc.sh

Acknowledgment

• This repository borrows heavily from beta-TCVAE. We thank them for open-sourcing their codes.

Bibtex

@article{kim2020semi,
  title={Semi-supervised Disentanglement with Independent Vector Variational Autoencoders},
  author={Kim, Bo-Kyeong and Park, Sungjin and Kim, Geonmin and Lee, Soo-Young},
  journal={arXiv preprint arXiv:2003.06581},
  year={2020}
}