Simpletransformers

Added

- Added support for XLMRoberta for multi-label tasks.

Added

- Added support for language model training (from scratch or fine-tuning).
- Added option to choose which metric should be used for early stopping.

Changed

- Switched to using the logging module over print for everything except running loss. (QuestionAnsweringModel - [travis-harper](https://github.com/travis-harper))
- Replaced more concatenated string paths with `os.path.join()` when creating `training_progress_scores.csv`.

Changed

- Replaced concatenated string paths with `os.path.join()` when creating `training_progress_scores.csv`. [sarthakTUM](https://github.com/sarthakTUM)

Fixed

- Fixed issue with cached eval features being used even when using `predict()` in `ClassificationModel` and `NERModel`.

Added

- Added classification report for NER for per-tag scores. [seo-95](https://github.com/seo-95)