For details see CHANGELOG.md.
- Transformer based Libinvent
- Prior registry to load internal priors more easily
- Strict validation of input configuration to ensure consistency
- Better JSON configuration file writing
- Metadata writing for all created RL and TL models
- Import functionality for scoring runmode
- Stages in staged learning can have their own diversity filters
- More memory efficient transformer models to handle larger numbers of input SMILES
- Additional (fragment) SMILES written to staged learning CSV
- TanimotoDistance renamed to TanimotoSimilarity
- Support for ChemProp multitask models: requires param.target_column
- Allow dot SMILES fragment separator for Lib/Linkinvent input
- Optional [scheduler] section for TL
- Example support script for RAScore
- A more complete RL/TL demo notebook
- Experimental data pipeline to preprocess SMILES for prior creation
- Various code improvements and fixes
For details see CHANGELOG.md.
- Upgrade to PyTorch 2.2: rerun
pip install -r requirements-linux-64.lock - 2 new notebooks demoing Reinvent with reinforcement learning and also transfer learning, includes TensorBoard visualisation and basic analysis
- New Linkinvent model code based on unified transformer
- New PubChem Mol2Mol prior
- Unknown token support for PubChem based transformer models
- New "device" config parameter to allow for explicit device e.g. "cuda:0"
- Optional SMILES randomization in every TL epoch for Reinvent
- Dataclass parameter validation for most scoring components
- Invalid SMILES are now written to the reinforcement learning CSV
- Code improvements and fixes
For details see CHANGELOG.md.
- Reworked TL code with added options and statistics
- Standardization can be switched off in TL (useful in new prior creation)
- Similarity calculation in TL made optional
- Updated script for empty classical Reinvent model creation
- Allow runs with only filter/penalty components
- Stable sigmoid functions
- Removed long chain check in SMILES processing
- Unified transformer code
- Filter apply to transformed scores
- Better memory handling in inception
- Better logging for Reinvent standardizer
- Inception filters for tokens not compatible with the prior
- Number of CPUs for TL (Mol2Mol pair generation) is 1 by default
- Tensorboard histogram bug fixed again
- Code improvements and fixes
For details see CHANGELOG.md.
- Scoring component MolVolume
- Scoring component for all 210 RDKit descriptors
- CSV and SMILES file reader for the scoring run mode
- Tobias Ploetz' (Merck) REINFORCE implementations of the DAP, MAULI and MASCOF RL reward strategies
- Number of CPUs can be specified for TL jobs: useful for Windows
- All prior models tagged with metadata and checked for integrity
- Code improvements and fixes
- Combined RL/CL (staged learning)
- New transformer model for molecule optimization
- Full integration of all generators with all algorithmic frameworks (RL, TL)
- Reworked scoring component utilizing a plugin mechanism for easy extension
- TOML configuration file format in addition to JSON: note that format is incompatible with release 3.x
- Major code rewrite
- Single repository for all code