[adam-sutton-1992]

Hihi,

Adding class weighting exposing some training params. Previously the loss was calculated via the CRF outputs which predicted output tokens. This improved performance over baseline loss by enforcing transitions between tokens to make sense (i.e. B->I or E prediction tokens only, O -> B or S prediction tokens, etc...). I tested class weights for each predicted class with the new stats module and making these changes results in performance increases. Here they are:

1. config.class_weights - A dict of how to weigh each label class towards the calculated loss. Generally speaking in this task making O lower increases recall at the cost of precision. Increasing B and S also encourages detecting the start of concepts more liberally. Defaults to a uniform weighting.
2. config.crf_loss_weight - Weigh the importance of the loss function that was previously the standard. This is crf outputs. Effectively you can increase/decrease the importance of the output being consistent.
3. config.learning_rate - The learning rate was set to 1e-5, without it being exposed to be edited nicely. This fixes that.
4. config.weight_decay - same as above.

This also comes with methods in transformer_ner which exposes these changes and nicely saves / loads the parameters.

The changes to class weights and loss weight can significantly improve performance on the NER Only task. Going from a best Recall of 0.9 to 0.96, where the vocab based nerrule based method with training can achieve about 9.75. This also has a higher precision compare to the vocab based ner method. Reaching about 0.4.

[mart-r]

Please resync with main branch. Hard to review if this includes changes that are (or may be) already merged. For instance, I can see that changes from #639 are presented in this PR as well.