This includes new features / bug fixes for semantic loss. The main changes are:
Argument Linking for the data module: Semantic loss and base loss automatically use the DataModule instance provided by the --data argument (instead of chebi100.yml), closes #23
Add softmax multiplier for semantic loss
Add PubChem K-means based dataset - clusters PubChem data into distinct parts based on similarity between fingerprints, excludes parts that are close to ChEBI
analyse_sem.py now selects the best checkpoint based on micro-f1 (instead of validation loss)
modularisation of analyse_sem.py, additional features (verbose output that shows individual violations, option to only calculate specific metrics, option to automatically remove violations)
This includes new features / bug fixes for semantic loss. The main changes are:
DataModuleinstance provided by the--dataargument (instead ofchebi100.yml), closes #23analyse_sem.pynow selects the best checkpoint based on micro-f1 (instead of validation loss)analyse_sem.py, additional features (verbose output that shows individual violations, option to only calculate specific metrics, option to automatically remove violations)