0.1.3

Important changes:

• Acceleration of mace and allegro models via the cuequivariance library from Nvidia.
• Added continuity testing to ensure backwards compatibility.
• Refactor and theme change of the documentation.
• Added neighborlist method as argument to several methods, positibility to use torch or nvidia implementations

0.1.2

This release includes several upgrades:

• Minor improvements in feature extraction for old PyG models.
• Improved compilation engine:
  • Avoids compilation in problematic cases.
  • Automatically retries using torch._dynamo.config.suppress_errors to detect non-compilable code paths and fall back to eager mode when needed.
• Modernization of the repository structure.

Full Changelog: v0.1.1...v0.1.2

0.1.1

This release includes some important upgrades:

• Allow for compilation of models for simulation.
• Increase the torch version to 2.6.
• Minor fixes regarding installation instructions.

0.1.0

The release incorporates very important new features:

• Added Allegro and So3krates as new possible architectures, as well as a complete new design of MACE based on the current version.
• Update the pytorch lightning version from 1.9.4 to 2.2.1. We include scripts/mlcg-adapt-training-yaml.py to adapt old training yaml scripts.
• Add automatic testing of the examples.

0.0.4

Small changes including:

• Split the mlcg/nn/prior.py monolith into several individual files for easier manipulation.
• Fixed a problem where PTSimulation was not restarting sims from the ckpt coordinates if the specialize_prior was not set.
• Ensured the examples/input_yamls files are compatible with pytorch lightning 1.9.4

0.0.3

This release incorporates some relevant changes such as:

• Improving the documentation about the units correlation in simulation.
• Adding a Quartic Prior
• Fixing a bug of the initial velocities in simulation when using units that were not kcal/mol

0.0.2

Version 0.0.2

This release has incorporated some relevant changes:

• Corrected some bugs occurring when only using the energy as output
• Added the option to exclude some neighbors from the Schnet computation.
• Refactored the examples folder

0.0.1

First Release

This is the code for training and simulating machine learned, coarse grained models for proteins.