codeaudit/jmvae

JMVAE

Code for reproducing results of our paper "Joint Multimodal Learning with Deep Generative Models"

We also developed a Python framework for deep generative models called Tars in Theano and Lasagne. As we implemented JMVAE as a model in Tars, please install this framework first before executing codes in this repository.

Installation

Run the following comamnds to install Tars (v0.0.2). Please make sure that you specify the version 0.0.2.

$ git clone https://github.com/masa-su/Tars.git -b v0.0.2
$ pip install -e Tars --process-dependency-links

When you execute the above commands, the following packages will be automatically installed in your environment:

• Theano
• Lasagne
• progressbar2
• matplotlib
• sklearn

Training models

Use main_jmvae_zero_z_x.py and main_jmvae_kl_z_x.py scripts to train JMVAE-zero and JMVAE-kl on the MNIST. If you want to train JMVAE-GAN on CelebA, first download the CelebA dataset. Then crop the images to 64×64, and put them to Tars/datasets directory before executing main_jmvaegan_kl_z_x.py.

Citation

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

@article{suzuki2016joint,
  title={Joint Multimodal Learning with Deep Generative Models},
  author={Suzuki, Masahiro and Nakayama, Kotaro and Matsuo, Yutaka},
  journal={arXiv preprint arXiv:1611.01891},
  year={2016}
}

and Tars.