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

Install (CPU version) using pip install -r requirements.txt

For GPU compatibility, replace the fourth line in requirements.txt with: https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-0.12.1-cp27-none-linux_x86_64.whl

To create the molecule datasets, call:

• python make_zinc_dataset_grammar.py
• python make_zinc_dataset_str.py

The equation dataset can be downloaded here: grammar, string

To train the molecule 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
• python train_zinc_str.py

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

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

• python encode_decode_zinc.py

The analogous file equation_vae.py can encode and decode equation strings. Run:

• python encode_decode_eq.py

The Bayesian optimization experiments use sparse Gaussian processes coded in theano.

We use a modified version of theano with a few add ons, e.g. to compute the log determinant of a positive definite matrix in a numerically stable manner. The modified version of theano can be insalled by going to the folder Theano-master and typing

• python setup.py install

The experiments with molecules require the rdkit library, which can be installed as described in http://www.rdkit.org/docs/Install.html.

The Bayesian optimization experiments can be replicated as follows:

1 - Generate the latent representations of molecules and equations. For this, go to the folders

molecule_optimization/latent_features_and_targets_grammar/

molecule_optimization/latent_features_and_targets_character/

equation_optimization/latent_features_and_targets_grammar/

equation_optimization/latent_features_and_targets_character/

and type

• python generate_latent_features_and_targets.py

2 - Go to the folders

molecule_optimization/simulation1/grammar/

molecule_optimization/simulation1/character/

equation_optimization/simulation1/grammar/

equation_optimization/simulation1/character/

and type

• nohup python run_bo.py &

Repeat this step for all the simulation folders (simulation2,...,simulation10). For speed, it is recommended to do this in a computer cluster in parallel.

2 - Extract the results by going to the folders

molecule_optimization/

equation_optimization/

and typing

• python get_final_results.py
• ./get_average_test_RMSE_LL.sh