Code for paper "Learning a Generative Model for Validity in Complex Discrete Structures" by David Janz, Jos van der Westhuizen, Brooks Paige, Matt Kusner and José Miguel Hernández-Lobato. Arxiv: https://arxiv.org/abs/1712.01664

Steps to reproduce results in table 4:

1. Install requirements. rdkit (chemical informatics, http://www.rdkit.org/docs/Install.html) can be tricky; we recommend using Anaconda. Note: the project uses pytorch 0.3.1, later versions are not backward compatible and will not run.

2. First prepare data by running python3 data data/zinc.smi

3. Generate samples from pretrained VAE for testing the grammar model. See optional) for training your own VAE. python3 vae --restore vae-pretrained --generate-samples

4. Train grammar_rnn model, the main paper contribution. This takes approximately ten minutes. python3 grammar_model

Results are displayed to terminal + plotted to tensorboard (default location /tmp/models)

Citation bibtex:

@article{janz2018learning, title={Learning a Generative Model for Validity in Complex Discrete Structures}, author={Janz, David and van der Westhuizen, Jos and Paige, Brooks and Kusner, Matt and Hern{'a}ndez-Lobato, Jos{'e} Miguel}, journal={International Conference on Learning Representations}, year={2018} }

Optional: Train your own character VAE for molecules. Enclosed implementation beats the original. Training takes a couple hours. If you get a cuda out of memory error, consider reducing --batch-size (defaults at 500). python3 vae