peterbjorgensen/msgnet

Tensorflow implementation of message passing neural networks for molecules and materials

msgnet's Issues

To clarify, the edge update functionality is only in `msgnet`, not in `schnetpack`, correct?

From README.md:

Tensorflow implementation of message passing neural networks for molecules and materials.
The framework implements the SchNet model and its extension with edge update network NMP-EDGE as well as the model used in Materials property prediction using symmetry-labeled graphs as atomic-position independent descriptors.

Currently the implementation does not enable training with forces, but this might be implemented in the future.
For a more full-fledged implementation of the SchNet model, see schnetpack.

The main difference between msgnet and schnetpack is that msgnet follows a message passing architecture and can therefore be more flexible in some cases, e.g. it can be used to train on graphs rather than on structures with full spatial information.

ASE databases scalar properties

Is there a reason that in get_matproj.py the scalar properties like delta_e and band_gap are saved as key_value_pairs and in get_qm9.py the 17 properties including tag and index are saved in the "data" dictionary, while they could also be stored as key_value_pairs, if I understand the ase.db documentation correctly? This seems to make the input pipeline a bit more complicated.