All notable changes to the project are documented in this file.

The format follows Common Changelog, and the project adheres to Semantic Versioning.

• Relax protobuf version requirement (#62)

• Fix hydrogen handling in scaffolds with explicit attachment points (#70)
• Avoid memory leaks and other tensorflow issues (#68)

• Add an option for decode to return the entire generation trace (#51)

• Reformat with black==23.1.0 and pin it in CI to avoid further unexpected updates (#50)

• Remove deprecated numpy types to make molecule_generation work with numpy>=1.24.0 (#49)
• Patch GetSSSR for compatibility with rdkit>=2022.09.1 (#58)

• Add support for fine-tuning a pretrained model on new data (#30)
• Add a __version__ attribute to make the package version easily accessible at runtime (#35)

• Drop the exact version requirements for python and tensorflow to support entire ranges of versions (#35)

• Drop unused GraphMultitaskModel which ended up in the open-source release by accident (#34)

• Make the inference server check the status of child processes every 10 seconds, so that it can exit gracefully in case of errors instead of hanging (#33)

• Add MoLeRGenerator, which uses the MoLeR decoder (without the encoder) as an autoregressive policy (#6)
• Add load_model_from_directory, which can load any model by automatically picking the right wrapper class (#24)
• Implement an option for encode to return not only the mean latent code but also the variance (#26)

• Improve how the MoLeR visualisers handle node selection steps (#10)
• Refactor how MoLeR keeps track of generation steps during decoding and include partial molecules in the step info classes (#27)

• Fix one-off errors in the latent-based visualisation mode (#10)
• Constrain protobuf version to avoid pulling in a breaking release (#25)

🌱 First public release, matching what was used for the original paper.

• Add full implementation of MoLeR as introduced in the paper
• Add reference implementation of CGVAE, not yet supported by the high-level model API