feat(medcat):CU-869cy3xa0 Improve training

[mart-r]

This PR does an overhaul to the training setup of MedCAT:

• It modifies the existing TrainableComponent protocol to also include a train_unsupervised method
  • And uses that over the "check config for train and run inference" unsupervised training
• It allows all components that follow the TrainableComponent protocol to be trained supervised
  • Previously only the linker was able to be trained in a supervised manner
• It provides a few utility methods to allow training and evaluating components individually
  • I.e dataset-aware components that will enable either training or evaluating only NER or Linker if/when required

Example code snippets:

1. When only training linker

with dataset_aware_component(cat, CoreComponentType.ner, DATASET):
    trainer.train_supervised_raw(DATASET, nepochs=1)

2. When only training NER

with dataset_aware_component(cat, CoreComponentType.linking, self.DATASET):
    trainer.train_unsupervised([doc['text'] for proj in self.DATASET['projects'] for doc in proj['documents']], nepochs=1)

3. When doing evaluation / stats one component at a time

with dataset_aware_component(cat, CoreComponentType.ner, self.DATASET):
    tps, fns, tps, cui_prec, cui_rec, cui_f1, cui_counts, examples = get_stats(
        cat, self.DATASET, do_print=False)

[adam-sutton-1992]

Does this mean we're still doing unsupervised on a per entity basis? I can't think of a case where in an unsupervised manner you would need the entity.

[mart-r]

This is supervised training still. This is within the add_and_train_concept method.

[adam-sutton-1992]

A few queries but I think it looks good. I might've missed these within the commits:

If you have two trainable components. is it possible to turn of training for one of them when running training methods? Do the dataset aware components serve that purpose?

And one more above^^^

[mart-r]

If you have two trainable components. is it possible to turn of training for one of them when running training methods? Do the dataset aware components serve that purpose?

The description already had 2 examples for this :)

The dataset aware implementation can serve that purpose. Because they replace the specific component with another one (which isn't trainable, but that's kind of irrelevant since it's a different component) for the duration of the context manager.

But I think what makes it unclear is that in the example I've given it a dataset, but realistically, you could provide an empty dataset for it, i.e like this:

# supervised
with dataset_aware_component(cat, CoreComponentType.ner, {"projects" : []}):
    trainer.train_supervised_raw(DATASET, nepochs=1)
# unsupervised
with dataset_aware_component(cat, CoreComponentType.ner, {"projects" : []}):
    trainer.train_unsupervisedsupervised(["list", "of", "texts'], nepochs=1)