diff_vae_gan

Adversarial Scaffold Regularization

The folder contains the training and testing scripts for variational junction tree encoder-decoder with adversarial training. The trained model is saved in models/ for all datasets. Sample translations of the test set are provided in results/ folder.

Preprocessing

Please make sure you first preprocess the paired translation data following (diff_vae/README.md). In addtion, we need to preprocess the "good" molecules in the target domain:

python ../scripts/preprocess.py --train ../data/qed/target.txt --mode single --ncpu 8
mkdir target-processed
mv tensor* target-processed

Training

Let's use QED task as a running example. Suppose your preprocessed translation pairs are saved in ../data/qed/processed and target domain molecules are saved in ../data/qed/target-processed. To train our model, run

mkdir -p newmodels/qed
python arae_train.py --train ../data/qed/processed/ --vocab ../data/qed/vocab.txt \ 
--ymols ../data/qed/target-processed/ --save_dir newmodels/qed \ 
--hidden_size 300 --rand_size 8 --epoch 10 --anneal_rate 0.8 | tee newmodels/qed/LOG

Here --ymols specifies the folder where target domain molecules are stored.

Validation

After training, you can evaluate different model checkpoints to select the one with the best performance:

bash val_scripts/valid_qed.sh models/qed 10

Testing

After finishing cross-validation, you can test the chosen model on the logp04 task by running

python decode.py --test ../data/qed/test.txt --vocab ../data/qed/vocab.txt --model models/qed/model.iter-3 | python ../scripts/qed_score > results.qed
python qed_analyze.py < results.qed

You can test our models on all four tasks by running

bash decode_and_eval.sh

or equivalently bash eval.sh if the decoded file is given. This should print the model performance as follows:

QED task:
Evaluated on 800 samples
success rate 0.615
DRD2 task:
Evaluated on 1000 samples
success rate 0.786

Note that the above results are slightly different from the reported value in the paper. In the paper, we run this evaluation with multiple random seeds and report the average performance. Here we run with just one random seed.