Better deploying and local word2vec computation with gensim

[AndreiBarsan]

Did a bit of engineering. My Euler queue is now pretty backlogged. :D Still WIP, as the code needs a little bit of cleanup.

• Added simple notebook for training word2vec embeddings locally on the Twitter dataset using the gensim library.
• Improved fabfile to also deploy to Euler seamlessly for training.
  • No support for auto-download of trained model from Euler yet;
  • No support for live TensorBoard on Euler, due to its batch system nature;
  • Compatible with Fabric3, which allows us to use Fabric with Python 3 (pip install Fabric3; see fabfile.py)
  • Preprocessing script now uses google-style flags.
  • Better parameterization for train_model.py as well.

[bernhard2202]

Awesome work, LGTM :) Can you maybe also post the output of the preprocessing with the user-trained embeddings?

[AndreiBarsan]

Thanks for reviewing! I'll post the results once I get them. It may take a while, so I may just have to wait until tomorrow.

[AndreiBarsan]

I got only 0.85600 on Kaggle, which isn't great. Slightly reduced learning rate (0.000075), trained embeddings for test data too, and reduced their dimensionality from 300 to 275. Running on kaggle now.