MolGAN-pytorch

🦑This repository is a Pytorch implementation of MolGAN: An implicit generative model for small molecular graphs.

Installation

Dependency

pip install -r requirements.txt

If you still run into no-such-package issues, you may just install the required packages by pip or conda.

Dataset Generation

# Download gdb Dataset and NP/SA scores
bash ./data/download_dataset.sh
# Generate QM9 Dataset
python ./data/sparse_molecular_dataset.py

Run

To train the MolGAN model, you could run the experiment with the following command:

python main_gan.py --lambda_wgan 0.05 --desc "lambda_wgan 0.05"

The lambda_wgan is a config parameter, which refers to the hyperparamer $\lambda$ that balances WGAN loss and RL loss in the original paper.

Some experiment results

lambda_wgan	Valid	Unique	Novel	Solubility
0(full RL)	78.21	93.58	97.51	0.30
0.01	80.01	95.90	97.79	0.31
0.05	77.04	97.19	99.24	0.33
0.1	69.51	96.40	98.93	0.32
0.25	68.90	95.00	98.88	0.33
0.5	73.81	96.31	100.00	0.31
0.75	82.60	95.95	99.50	0.33
1(no RL)	81.29	95.65	97.98	0.31