fabiannagel / schnetkit Goto Github PK

This is a fork of schnetkit for benchmarking SchNetPack's neighbor list machinery.

EXPERIMENTAL/IN DEVELOPMENT DO NOT USE

This is an early draft of some infrastructure built around schnetpack. In particular, it allows:

• Saving/loading models from disk
• Initialising models from a .yaml-based description of the architecture
• Efficient ase calculator for SchNet models with support for a larger effective cutoff (i.e. avoiding to recompute the neighborlists on every step)

As a result of this, the schnetkit calculator is compatible with FHI-vibes. (Currently, this functionality is untested.)

Future versions will support:

• Restarteable training with automatic re-submission in slurm
• Training configurable with input files
• Efficient dataset class that pre-computes everything

This is a part of the infrastructure for the gknet project, and will be completed as the project nears publication.