Feature/fix gec predictor

[Frost45]

In this PR, code changes have been made to the following files:

1. datareader.py
2. gec_predictor.py
3. seq2labels_model.py
4. test_seq2labels.py
5. regression_tests/test_regression_data_predictor.py

Please review the above files.

I've added additional test cases to:

1. test_gec_model.py
2. test_gec_predictor.py

I've left in the comments that I added to gec_model.py for understanding what was happening.

I have also rebased the master branch onto this branch.

How to test:

• pytest
• python regression_tests/test_regression_data_predictor.py

[desilinguist]

Thanks for figuring this out! I had some efficiency and readability suggestions.

[desilinguist]

Thanks for addressing my comments! LGTM!

[ksteimel]

Thanks for fixing these discrepancies! I'm sure this wasn't easy to track down!
I do wish stuff like appending $START was done in the dataset reader's text_to_instance method. That'd be the more allennlp-y way to handle it. However, this would break the existing gec_model.py code which is expecting to have to add $START so that doesn't really seem possible right now.

[Frost45]

Thanks for fixing these discrepancies! I'm sure this wasn't easy to track down! I do wish stuff like appending $START was done in the dataset reader's text_to_instance method. That'd be the more allennlp-y way to handle it. However, this would break the existing gec_model.py code which is expecting to have to add $START so that doesn't really seem possible right now.

@ksteimel When we add ensemble prediction to the predictor, we can get rid of GEC_model entirely and make this change.