v0.3.0

Code for training and validating the SO3krates model from the paper https://arxiv.org/abs/2309.15126.

Below changes wrt version 0.2.0 are listed.

What's Changed

• Orbax migration by @thorben-frank in #9
• Checkpoint asynchronous dispatch by @thorben-frank in #10
• allow to pass individual train and valid files by @thorben-frank in #12
• Remove jax-md and replace with glp by @sirmarcel in #13
• Repeated degree support by @thorben-frank in #14
• Energy repulsion by @thorben-frank in #15
• Restart training from checkpoint by @thorben-frank in #18
• Rework ase calculator by @thorben-frank in #19
• new dataloader by @kabylda in #21
• integrate ASE->jraph data loader for sparse neighborlists by @thorben-frank in #22
• add capacity_multiplier to create_from_ckpt_dir method by @thorben-frank in #23

New Contributors

• @sirmarcel made their first contribution in #13
• @kabylda made their first contribution in #21

Full Changelog: v0.2.0...v0.3.0

v0.2.0 Last version with depracated FLAX checkpointing

This version is the last mlff version that can deal with the depracated flax checkpointing system. Explanation see below.

With v0.7.0, flax deparactes their internal checkpointing system and moves to orbax. This has the consequence, that old parameter checkpoints can not be restored with the newest flax version anymore and, unfortunately, breaks backward compability. Starting from v0.3.0 mlff moves to orbax, thus older flax checkpoint are not restorable anymore.