sdvae

Syntax-Directed Variational Autoencoder for Structured Data (https://arxiv.org/abs/1802.08786)

1. Download the data and pretrained model

Use the following dropbox link:

https://www.dropbox.com/sh/621ufmvqgg5h2d8/AAARWPpuADNfPx8eu9E8y-rha?dl=0

Put everything under the 'dropbox' folder, or create a symbolic link with name 'dropbox':

ln -s /path/to/your/downloaded/files dropbox

Finally the folder structure should look like this:

sdvae (project root)
|__  README.md
|__  mol_vae
|__  prog_vae
|__  dropbox
|__  |__ data
|    |__ results
|    |__ context_free_grammars
|......

2. install dependencies and build c++ backend

The current code depends on pytorch 0.3.1. Most of the python dependencies can be installed by pip. However, the bayesian optimizaiton depends on a customized build of Theano. Please follow the instruction in GrammarVAE (https://github.com/mkusner/grammarVAE):

below we will use mol_vae as the illustration for training/evaluation. The prog_vae works similarly.

3. Dataset pre-processing

Before training/evaluation, we need to cook the raw txt dataset. We use the mol_vae as illustration:

cd mol_vae/data_processing
./run_data.sh
./run_cfg_dump.sh

The above two scripts will compile the txt data into binary file and cfg dump, correspondingly.

4. Training

To train the model using GPU, run the following commands. You may also want to modify the parameters in the training script.

cd mol_vae/pytorch_train
./run_train.sh

The pretrained models are available under the dropbox folder, dropbox/results.

5. Evaluation

Before evaluation, we need to first dump the latent encodings of programs/molecules:

cd mol_vae/pytorch_eval
./run_feature_dump.sh

To test the reconstruction, or sample from prior, please see the corresponding scripts under the same folder.

5.1 Bayesian Optimization

To optimize the molecule property, run the bayesian optimization:

cd mol_vae/mol_optimization
./run_bo.sh

After that, use the script get_final_results.py to collect the results. We use the same evaluation protocol as in GrammarVAE(https://github.com/mkusner/grammarVAE).

The results reported in the paper can be found under dropbox/results/zinc/bo. If you use the same random seeds, then the exact same results should be expected.

5.2 Sparse Gaussian Regression

To test the regression performance using the latent embeddings of molecules/programs:

cd mol_vae/sparse_gp_regression
./run_regression.sh

Again, the 10 runs with different random seeds are reported, under dropbox/results/zinc/sgp

5.3 Visualization of Latent Space

To interpolate the latent space, do the following:

cd mol_vae/visualize
./run_2dvis.sh

You may want to tune the gap, number of grids, etc., to see some reasonable visualization results.

Reference

@article{dai2018syntax,
  title={Syntax-Directed Variational Autoencoder for Structured Data},
  author={Dai, Hanjun and Tian, Yingtao and Dai, Bo and Skiena, Steven and Song, Le},
  journal={arXiv preprint arXiv:1802.08786},
  year={2018}
}