This is a PyTorch implementation of "Variational Autoencoders with Jointly Optimized Latent Dependency Structure" which appeared in ICLR '19.

Clone the repository

git clone https://github.com/ys1998/vae-latent-structure.git
cd vae-latent-structure/

Ensure that the requirements mentioned here are met.

# train
python train.py --config config.json
# resume from checkpoint
python train.py --resume <path_to_checkpoint>
# using multiple GPUs (equivalent to "CUDA_VISIBLE_DEVICES=2,3 python train.py -c config.py")
python train.py --device 2,3 -c config.json 
# test
python test.py --resume <path_to_checkpoint>
# visualization
tensorboard --logdir <path_to_log_dir>

You can find a detailed description of the folder structure here. The code is distributed across several branches, as described below

1. master: original implementation (GraphVAE)
2. LSTM: replacing top-down inference using a recurrent network (GraphLSTMVAE)
3. vrnn: extension of GraphVAE to sequential data (RecurrentGraphVAE)

All models have been described in the model folder in the respective branch. Please look at our report for more details on our variants and results.

All config files are specified in JSON format as described here. Model-specific parameters/options are provided via the arch field. The config.json files in each branch already contain default parameters for the respective model.

We would like to thank the authors of the PyTorch template https://github.com/victoresque/pytorch-template which served as a starting point for our project.

Jiawei He and Yu Gong and Joseph Marino and Greg Mori and Andreas Lehrmann. Variational Autoencoders with Jointly Optimized Latent Dependency Structure, ICLR 2019. URL https://openreview.net/forum?id=SJgsCjCqt7