Multi-view Chemical Variational Autoencoder
Installation
A conda environment is necessary for installation. To create the enviroment, run the following command:
conda env create -f environment.yml
source activate multiview_chemvae
Jupyter notebook is also required to run the *.ipynb examples in the experiments directory.
Creating datasets
To create the datasets for training, run
python make_char_dataset.py
python make_grammar_dataset.py
python make_features_dataset.py
The datasets will be named char_dataset.h5, grammar_dataset.h5, and features_dataset.h5 respectively.
Training
To train a model, run
python train.py --model_type <type>
where <type> is either "Grammar" or "Character".
Additional flags include:
--two_tower: specifies whether to train a two-tower model or not--load_model: load weights from an existing model--latent_dim: specifies dimensionality of latent space--epochs: specifies number of training epochs
For example:
python train.py --model_type=Grammar --two_tower --latent_dim=128 --epochs=100
Experiments
The experiments directory contains the following Jupyter notebooks for evaluating the performance of the models:
-
Prior validity:
experiments/prior_validity.ipynb -
Reconstruction accuracy:
experiments/reconstruction_accuracy.ipynb -
Property prediction:
experiments/property_prediction.ipynb