GitLab

Added top proportional and setting files as well as al setting for mutil rc that sample same number of samples as for movies

Added stratified sampling for active_learners and add config files for movies and multirc for hxl and hxlp

Now the model parameters are actually reused.

* Fixed movies_expred.json: classifier epoch = 10 (was 2)
* Added params for  *al_auxltu_one_step_reuse experiments

* Before always the model after the last epoch has been chosen now you can select if you want to restore the best model based on the validation loss by setting the config value "best_model_selection" to True

implemented MaximumNormalizedTrueLogProbability an oracle al method for comparison that know the true class of the tokens and uses the MNLP criterion based on
    the probability of the true class