This repository holds the code for the Keras layers implementation of Behler's Atom-Centered Symmetry Function, in the context of the following paper:
Profitt, T.A. and Pearson, J.K., 2019. A shared-weight neural network architecture for predicting molecular properties. Physical Chemistry Chemical Physics, 21(47), pp.26175-26183.
The easiest way to use the layers is to install them as the package using the normal package management methods for Python. The layers are not yet released as a Python package under PyPI, so one will have to clone this repository and install using the files themselves.
git clone https://github.com/UPEIChemistry/atomic-images.gitThen install by pointing pip to the directory:
pip install /path/to/atomic-images/Once installed, the layers should be importable as normal:
from atomic_images import layersIn the examples directory, one can find examples of using the layers for both numpy operations as well as Keras models. Consult the source code and docstrings in atomic_images/layers.py for more information regarding input and output shapes and parameters.
I intend to write a full script to demonstrate usage on the QM9 in the future, but in the current state of the repository, the layers should be usable to build ASCF models in Keras.
Contributions are welcome. You can fork the repository and submit a pull request to have documentation or code changed. Submit issues or pull requests for reporting bugs or discussing potential changes.