Adding the code which helped me place in the top 1% of the kaggle tranfer learning challenge.
It has preprocessing steps to get only relavent information from the text, standardizing it by looking for POS tags, lematizing it.
It then goes uses the 6b glove embeddings for the transfer learning part, then learns the model on a embeddings + cnn + lstm network and uses hamming loss to pick the right cutoffs per label.
I haven't added the glove embeddings or the test, train data since the are above the github threshold, feel free to create a issue in case you are unable to find them.
Happy Kaggling!