RandomNets is an efficient, vectorized solution for creating and training implicit ensemble feed-forward neural networks, designed to enhance prediction robustness and performance. By introducing input feature masking, RandomNets generates diverse predictions within a single model architecture, making it an attractive alternative to standard feed-forward neural networks. The model is especially suited for molecular property prediction tasks and has demonstrated superior performance across a wide range of bioactivity datasets.
- Implicit Ensembles: Achieves ensemble performance with minimal computational overhead by leveraging vectorization.
- Feature Masking: Adds diversity to predictions, optimizing performance with deterministic inference.
- Scalability: Training time grows sublinearly with ensemble size, enabling efficient scaling to larger ensembles.
- Open Source: Provided under an LGPL license for easy adoption and adaptation.
To install the RandomNets package, you can use pip with the following command:
pip install git+https://github.com/EBjerrum/randomnetsAfter installation, you can quickly set up and train a RandomNets model for your molecular property prediction tasks using the Pytorch Lightning compatible FpsDataModule and RandomNets model.
I've put together a small tutorial in this Notebook
Refer to PyTorch lighning documentation for training and inference.
If you use RandomNets in your work, please cite the associated preprint available on ChemRxiv:
Bjerrum, E. J. (2024). RandomNets Improve Neural Network Regression Performance via Implicit Ensembling. ChemRxiv. https://chemrxiv.org/engage/chemrxiv/article-details/67656cfa81d2151a02603f48
The paper contains details on approach, implementation, testing and results from 133 bioactivity datasets.
Contributions, bug reports, and suggestions are welcome! Please open an issue, or reach out (I'm googlable) or submit a pull request to improve the project. I recommend reaching out first to avoid duplication of work.
This project is licensed under the LGPL license. See the LICENSE file for details.
For commercial support or consultancy, please contact Cheminformania Consulting