README.md

Grammar Variational Autoencoder

This repository contains training and sampling code for the paper: Grammar Variational Autoencoder.

Creating datasets

To create the molecule datasets, call in this directory:

unzip ../rockyou-processed.zip -d data/
python make_zinc_dataset_grammar.py

Training

To train the models, call:

python train_zinc.py % the grammar model
python train_zinc.py --latent_dim=2 --epochs=50` % train a model with a 2D latent space and 50 epochs

Sampling

The file molecule_vae.py can be used to encode and decode SMILES strings. For a demo run:

python encode_decode_zinc.py
python encode_decode_zinc.py --latent_dim=2 --epochs=50 --num_samples=200 % sample 200 random strings from the model trained with a 2D latent space and 50 epochs

Other details

• Model is defined in models/model_zinc.py. Modify _buildEncoder and _encoderMeanVar in a similar manner to change the model i.e. add same additonal layers to both functions.

• The folder results/ contains some trained models which can be samples using the above commands. Note that the arguments need to be fed appropriately to the files.

• Some samples already generated from the modles are in samples/.