Various features

[sfluegel05]

This branch includes several new features, including:

• stratification of train-val-test split (ensures near-equal distribution of classes between data splits)
• Inner cross-validation
• specifying chebi version (e.g. v227 instead of default v200), also specfying separate chebi versions for train and test set
• using DeepSMILES and SELFIES instead of SMILES for pretraining on Pubchem and classification on ChEBI
• fixed calculation of macro-f1 during training
• added functions for model evaluation
• support for wandb logging
• best checkpoints are saved based on maximum micro-f1 score
• support for chebai_graph module
• command predict_from_file to predict SMILES strings directly

Compatibility issues with current dev version:

• structure of data-folder changed to include chebi-version, keeping existing datasets requires moving data manually, e.g.
  • old: data/ChEBI100/processed/smiles_token/test.pt
  • new: data/ChEBI100/chebi_v200/processed/smiles_token/test.pt
• some config files might have been changed / renamed
• content and location of tokens.txt have changed (leads to different encoding of tokens during dataset creation)
• might have some dependencies not included in requirements.txt

Open problems:

• Cross-validation does not work as intended -> reimplemented, works now
• uses wandb by default -> separate configs for wandb logger and csv logger
• macro-f1 results are wrong -> fixed
• Lint checks fail -> fixed after reformating