Rewrite

[ieaves]

I re-wrote a substantial chunk of the code-base for my own purposes and am opening this PR primarily in case there were any components you would like to reuse - please feel free to close it if not.

Primary changes

• An sklearn API compatible transformer, model, and pipeline object for the default model
• Performance improvements to pre-processing
• Various bug fixes the largest of which covers missing value handling
• reorganized package (dependencies need to be tested and a few other tweaks but this is quite close to being packaged and thrown up onto pypi if you were so inclined)

A few notes

1. Training code requires a few tweaks to work with the rest of the modified codebase
2. The preprocessing code used to generate training data has broadly broken missing value handling. The root problem is the application of .astype(str) to training data as this maps np.nan -> 'nan'. The downstream assumptions producing bag_of_$(x), paragraph vectors, and word_embeddings are all broken in the presence of missing values. The default sherlock model with correctly handled missing values now fails for sequences with missing values and could benefit from re-training.

Thanks for this great resource!