This version introduces some major changes. Most notably,
- ELECTRA tokenization changed (#169) to a static tokenizer (the old tokenizer is still available, but not the default anymore)
- A fresh inference pipeline (#148)
- ChEBI 2.0 support and SDF as the new input format instead of SMILES (the ELECTRA model still runs on SMILES, but these are generated from the ChEBI SDF file and canonicalized instead of taking the SMILES directly from ChEBI) (#147)
All changes
- Adding more metrics to evaluation by @schnamo in #149
- Fix review comments: type hints, error handling, and documentation by @Copilot in #152
- [Feature]: Generalize Prediction pipeline for Lightning CLI models by @aditya0by0 in #148
- fix checkpoint loading for electra, return attentions by @sfluegel05 in #157
- Update to ChEBI 2.0 & SDF data by @sfluegel05 in #147
- Use chebi-utils library by @sfluegel05 in #158
- Fix PubChem pretraining by @sfluegel05 in #159
- fix splits loading / inference tutorial by @sfluegel05 in #161
- chaning learning rate back to default of 1e-3 by @schnamo in #156
- fixes for tox21 dataset by @schnamo in #163
- Move jsonargparse to main dependencies, lazy-load sklearn/scipy in PubChem by @Copilot in #153
- add chebiFromList dataset by @sfluegel05 in #164
- Make PubChem datasets dynamic / Remove unsupported PubChem classes by @sfluegel05 in #167
- Static Tokenizer by @sfluegel05 in #169
Full Changelog: v1.2.0...v1.3.0
Contributors
schnamo, sfluegel05, and aditya0by0