Feature/automatic refinement transfer learning

[finnkap]

This Pull requests unites all the work done by the DLOmix group of the Advanced Practical course (Sylvie Baier, Lina-Liv Willruth, Julian Poschenrieder and Finn Kapitza).

New features are:

• interface that can be used from the Oktoberfest package:
  • downloading the baseline model from the repository
  • processing of a dataset
  • DLOmix model loading and saving
• Automatic refinement and transfer learning pipeline (possible with wandb integration)
• Report generation for the automatic refinement and transfer learning (works without wandb)
• functionality to change input and output layers of the network (allows for partial keeping of the weights)
• functionality to freeze and release specific parts of the model
• custom callbacks for logging and monitoring the automatic training process
• test ratio in dataset can be specified
• inference only dataset (no label column needed)
• dataset can now store an ordered list of ion types which appear in the data
• added optional advanced splitting to the dataset class in dlomix (based on PROSPECT splitting)

Fix:

• fix in parsing of sequences, where the alphabet does not match the data tokens
• regex for sequence processor was extended to also capture starting modifications from sequences without specific n- and c-terminal modifiers
• made custom loss function serializable for easy saving and loading of models