Support for Multitask Learning in Flair

[alanakbik]

This PR adds support for multitask learning in Flair (closes #2508 and closes #1260) with hopefully a simple syntax to define multiple tasks that share parts of the model.

The most common part to share is the transformer, which you might want to fine-tune across several tasks. Instantiate a transformer embedding and pass it to two separate models that you instantiate as before:

# --- Embeddings that are shared by both models --- #
shared_embedding = TransformerDocumentEmbeddings("distilbert-base-uncased", fine_tune=True)

# --- Task 1: Sentiment Analysis (5-class) --- #
corpus_1 = SENTEVAL_SST_GRANULAR()

model_1 = TextClassifier(shared_embedding,
                         label_dictionary=corpus_1.make_label_dictionary("class"),
                         label_type="class")

# -- Task 2: Binary Sentiment Analysis on Customer Reviews -- #
corpus_2 = SENTEVAL_CR()

model_2 = TextClassifier(shared_embedding,
                         label_dictionary=corpus_2.make_label_dictionary("sentiment"),
                         label_type="sentiment",
                         )

# -- Define mapping (which tagger should train on which model) -- #
multitask_model, multicorpus = make_multitask_model_and_corpus(
    [
        (model_1, corpus_1),
        (model_2, corpus_2),
    ]
)

# -- Create model trainer and train -- #
trainer = ModelTrainer(multitask_model, multicorpus)
trainer.fine_tune(f"resources/taggers/multitask_test")

The mapping part here defines which tagger should be trained on which corpus. By calling make_multitask_model_and_corpus with a mapping, you get a corpus and model object that you can train as before.

In addition, this PR makes some experimental changes that will likely be adapted further:

• adds support for gradient reversal to all models inheriting from DefaultClassifier
• changes loss reduction from 'mean' to 'sum' in DefaultClassifier
• removes the loss calculation in the final test in training

Edit: In addition, this PR removes some model classes that were very beta: the DependencyParser, the DistancePredictor and the SimilarityLearner. These may get added back when the model interfaces are finalized.