1. normalization = False works smoothly, except usually this leads to a poorly performing model.
2. python -m readability-transformers.reset_cache
3. Improvements for ResnetFC with more standard approaches
4. Added regression & classification head for these tasks
5. checks if train_metric actually requires the standard_errors parameter, whereas previously it just assumed it did and threw an error when using a train_metric that doesnt accept standard_errors