QFCMolGAN

This code is built on top of the PyTorch implementation for the article Exploring the Advantages of Quantum Generative Adversarial Networks in Generative Chemistry available here (from the main branch).

The primary addition, apart from considerable cleaning-up, is a PyTorch-implementation of a simulation of the QFC system for the input generator. Several input encodings are available, mostly for experimenting.

The quantum circuits are implemented in the models/q_generator module, implemented as PyTorch modules so that it works directly with backpropagation for simplicity and simulation speed. In a real experiment, this is of course not possible; the training of the quantum circuit should probably be split into a separate training step (interspersed with training steps of the generator and discriminator).

Environment

The environment can be installed via conda, mamba, or similar (commands are the same for both conda and mamba):

mamba env create -f environment.yml

You can then activate the environment:

mamba activate qfcmolgan

Download GDB-9 Dataset

Simply run a bash script in the data directory and the GDB-9 dataset will be downloaded and unzipped automatically together with the required files to compute the NP and SA scores.

bash data/download_dataset.sh

The QM9 dataset is located in the data directory as well.

Feel free to use it.

Data Preprocessing

Simply run the python script within the data directory.

You need to comment or uncomment some lines of code in the main function.

python data/sparse_molecular_dataset.py

Running the training script

The training is performed by running the main script

python main.py

You can define the training parameters within the training block of the main function in main.py.

Testing Phase

Simply run the same command to test the system.

You need to comment the training section and uncomment the testing section in the main function of main.py, and then run it as before.

Others

The results folder stores the log files, trained models, pretrained quantum circuits, and the testing results. If config.use_external_logger is enabled, the script will try to connect to a wand account and log the training progress there.

Citation

@misc{https://doi.org/10.48550/arxiv.2210.16823,
  doi = {10.48550/ARXIV.2210.16823},

  url = {https://arxiv.org/abs/2210.16823},

  author = {Kao, Po-Yu and Yang, Ya-Chu and Chiang, Wei-Yin and Hsiao, Jen-Yueh and Cao, Yudong and Aliper, Alex and Ren, Feng and Aspuru-Guzik, Alan and Zhavoronkov, Alex and Hsieh, Min-Hsiu and Lin, Yen-Chu},

  keywords = {Quantum Physics (quant-ph), FOS: Physical sciences, FOS: Physical sciences},

  title = {Exploring the Advantages of Quantum Generative Adversarial Networks in Generative Chemistry},

  publisher = {arXiv},

  year = {2022},

  copyright = {Creative Commons Attribution Non Commercial No Derivatives 4.0 International}
}

Credits

This repository was originally imported from: pykao/QuantumMolGAN-PyTorch , which refers to the following repositories:

• nicola-decao/MolGAN
• ZhenyueQin/Implementation-MolGAN-PyTorch
• jundeli/quantum-gan